Natural language serves as a common and straightforward control signal for humans to interact seamlessly with machines. Recognizing the importance of this interface, the machine learning community is investing considerable effort in generating data that is semantically coherent with textual instructions. While strides have been made in text-to-data generation spanning image editing, audio synthesis, video creation, and beyond, low-resource areas characterized by expensive annotations or complex data structures, such as molecules, motion dynamics, and time series, often lack textual labels. This deficiency impedes supervised learning, thereby constraining the application of advanced generative models for text-to-data tasks. In response to these challenges in the low-resource scenario, we propose Text2Data, a novel approach that utilizes unlabeled data to understand the underlying data distribution through an unsupervised diffusion model. Subsequently, it undergoes controllable finetuning via a novel constraint optimization-based learning objective that ensures controllability and effectively counteracts catastrophic forgetting. Comprehensive experiments demonstrate that Text2Data is able to achieve enhanced performance regarding controllability across various modalities, including molecules, motions and time series, when compared to existing baselines.

Programming based approaches to reasoning tasks have substantially expanded the types of questions models can answer about visual scenes. Yet on benchmark visual reasoning data, when models answer correctly, they produce incorrect programs 33% of the time. These models are often right for the wrong reasons and risk unexpected failures on new data. Unit tests play a foundational role in ensuring code correctness and could be used to repair such failures. We propose Visual Unit Testing (ViUniT), a framework to improve the reliability of visual programs by automatically generating unit tests. In our framework, a unit test is represented as a novel image and answer pair meant to verify the logical correctness of a program produced for a given query. Our method leverages a language model to create unit tests in the form of image descriptions and expected answers, followed by image synthesis to produce corresponding images. We conduct a comprehensive analysis of what constitutes an effective visual unit test suite, exploring unit test generation, sampling strategies, image generation methods, and varying the number of programs and unit tests. Additionally, we introduce four applications of visual unit tests: best program selection, answer refusal, re-prompting, and unsupervised reward formulations for reinforcement learning. Experiments with two models across three datasets in visual question answering and image-text matching demonstrate that ViUniT improves model performance by 11.4 points in accuracy. Notably, it enables 7B open-source language models to outperform gpt-4o-mini in visual program generation by an average of 7.7 points and reduces the occurrence of programs that are correct for the wrong reasons by 40%.

Zixian Ma, Jianguo Zhang, Zhiwei Liu, Jieyu Zhang, Juntao Tan, Manli Shu, Juan Carlos Niebles, Shelby Heinecke, Huan Wang, Caiming Xiong, Ranjay Krishna, Silvio Savarese. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Artemis Panagopoulou, Le Xue, Honglu Zhou, Silvio Savarese, Ran Xu, Caiming Xiong, Chris Callison-Burch, Mark Yatskar, Juan Carlos Niebles. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Jianguo Zhang, Thai Quoc Hoang, Ming Zhu, Zuxin Liu, Shiyu Wang, Tulika Manoj Awalgaonkar, Akshara Prabhakar, Haolin Chen, Weiran Yao, Zhiwei Liu, Juntao Tan, Juan Carlos Niebles, Shelby Heinecke, Huan Wang, Silvio Savarese, Caiming Xiong. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Large language model (LLM) agents have shown great potential in solving real-world software engineering (SWE) problems. The most advanced open-source SWE agent can resolve over 27% of real GitHub issues in SWE-Bench Lite. However, these sophisticated agent frameworks exhibit varying strengths, excelling in certain tasks while underperforming in others. To fully harness the diversity of these agents, we propose DEI (Diversity Empowered Intelligence), a framework that leverages their unique expertise. DEI functions as a meta-module atop existing SWE agent frameworks, managing agent collectives for enhanced problem-solving. Experimental results show that a DEI-guided committee of agents is able to surpass the best individual agent's performance by a large margin. For instance, a group of open-source SWE agents, with a maximum individual resolve rate of 27.3% on SWE-Bench Lite, can achieve a 34.3% resolve rate with DEI, making a 25% improvement and beating most closed-source solutions. Our best-performing group excels with a 55% resolve rate, securing the highest ranking on SWE-Bench Lite. Our findings contribute to the growing body of research on collaborative AI systems and their potential to solve complex software engineering challenges.

Achieving effective unified pretraining on large time series corpora remains an open challenge in developing time series foundation models. Existing methods, such as Moirai, introduce multiple projection layers for time series of different frequencies to account for high data heterogeneity. We identify major drawbacks to this human-imposed frequency-level model specialization. First, frequency is not a reliable indicator for grouping pretraining data. Second, time series can display varied distributions even within a short window. Frequency-level specialization overlooks the diversity at this granularity. To address these issues, this paper introduces Moirai-MoE, excluding human-defined data groupings while delegating the modeling of diverse time series patterns to the sparse mixture of experts (MoE) within Transformers. With this design, Moirai-MoE eliminates reliance on heuristics and enables automatic token-level specialization. Extensive evaluations on 39 datasets demonstrate the superiority of Moirai-MoE over state-of-the-art foundation models. This study also conducts comprehensive model analyses to explore the inner workings of time series MoE foundation models.

We introduce Reward-Guided Speculative Decoding (RSD), a novel framework aimed at improving the efficiency of inference in large language models (LLMs). RSD synergistically combines a lightweight draft model with a more powerful target model, incorporating a controlled bias to prioritize high-reward outputs, in contrast to existing speculative decoding methods that enforce strict unbiasedness. RSD employs a process reward model to evaluate intermediate decoding steps and dynamically decide whether to invoke the target model, optimizing the trade-off between computational cost and output quality. We theoretically demonstrate that a threshold-based mixture strategy achieves an optimal balance between resource utilization and performance. Extensive evaluations on challenging reasoning benchmarks, including Olympiad-level tasks, show that RSD delivers significant efficiency gains against decoding with the target model only (up to 4.4X fewer FLOPs), while achieving significant better accuracy than parallel decoding method on average (up to +3.5). These results highlight RSD as a robust and cost-effective approach for deploying LLMs in resource-intensive scenarios.

Kung-Hsiang Huang, Akshara Prabhakar, Sidharth Dhawan, Yixin Mao, Huan Wang, Silvio Savarese, Caiming Xiong, Philippe Laban, Chien-Sheng Wu. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Jierui Li, Hung Le, Yingbo Zhou, Caiming Xiong, Silvio Savarese, Doyen Sahoo. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Jianguo Zhang, Tian Lan, Ming Zhu, Zuxin Liu, Thai Quoc Hoang, Shirley Kokane, Weiran Yao, Juntao Tan, Akshara Prabhakar, Haolin Chen, Zhiwei Liu, Yihao Feng, Tulika Manoj Awalgaonkar, Rithesh R N, Zeyuan Chen, Ran Xu, Juan Carlos Niebles, Shelby Heinecke, Huan Wang, Silvio Savarese, Caiming Xiong. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Training effective AI agents for multi-turn interactions requires high-quality data that captures realistic human-agent dynamics, yet such data is scarce and expensive to collect manually. We introduce APIGen-MT, a two-phase framework that generates verifiable and diverse multi-turn agent data. In the first phase, our agentic pipeline produces detailed task blueprints with ground-truth actions, leveraging a committee of LLM reviewers and iterative feedback loops. These blueprints are then transformed into complete interaction trajectories through simulated human-agent interplay. We train a family of models---the xLAM-2-fc-r series with sizes ranging from 1B to 70B parameters. Our models outperform frontier models such as GPT-4o and Claude 3.5 on $\tau$-bench and BFCL benchmarks, with the smaller models surpassing their larger counterparts, particularly in multi-turn settings, while maintaining superior consistency across multiple trials. Comprehensive experiments demonstrate that our verified blueprint-to-details approach yields high-quality training data, enabling the development of more reliable, efficient, and capable agents. We open-source both the synthetic data collected and the trained xLAM-2-fc-r models to advance research in AI agents.Dataset: https://huggingface.co/datasets/Salesforce/APIGen-MT-5k & Models: https://huggingface.co/collections/Salesforce/xlam-2-67ef5be12949d8dcdae354c4

We present DyMU, an efficient, training-free framework that dynamically reduces the computational burden of vision-language models (VLMs) while maintaining high task performance. Our approach comprises two key components. First, Dynamic Token Merging (DToMe) reduces the number of visual token embeddings by merging similar tokens based on image complexity, addressing the inherent inefficiency of fixed-length outputs in vision transformers. Second, Virtual Token Unmerging (VTU) simulates the expected token sequence for large language models (LLMs) by efficiently reconstructing the attention dynamics of a full sequence, thus preserving the downstream performance without additional fine-tuning. Unlike previous approaches, our method dynamically determines token length based on the *image content*—not just resolution—and operates completely training-free, making it readily applicable to most state-of-the-art VLM architectures. Extensive experiments on image and video understanding tasks, demonstrate that DyMU can reduce the average visual token count by 32%-85% while achieving comparable performance to full-length models, across diverse VLM architectures. Furthermore, qualitative analyses show that the adaptive token reduction from DToMe aligns well with human perception and enables users to better control computational costs through flexible integration with additional vision tools and models.

Incorporating human feedback has been shown to be crucial to align text generated by large language models to human preferences. We hypothesize that state-of-the-art instructional image editing models, where outputs are generated based on an input image and an editing instruction, could similarly benefit from human feedback, as their outputs may not adhere to the correct instructions and preferences of users. In this paper, we present a novel framework to harness human feedback for instructional visual editing (HIVE). Specifically, we collect human feedback on the edited images and learn a reward function to capture the underlying user preferences. We then introduce scalable diffusion model fine-tuning methods that can incorporate human preferences based on the estimated reward. Besides, to mitigate the bias brought by the limitation of data, we contribute a new 1.1M training dataset, a 3.6K reward dataset for rewards learning, and a 1 K evaluation dataset to boost the performance of instructional image editing. We conduct extensive empirical experiments quantitatively and qualitatively, showing that HIVE is favored over previous state-of-the-art instructional image editing approaches by a large margin.

Recent advancements in multimodal pretraining have shown promising efficacy in 3D representation learning by aligning multimodal features across 3D shapes, their 2D counterparts, and language descriptions. However, the methods used by existing frameworks to curate such multimodal data, in particular language descriptions for 3D shapes, are not scalable, and the collected language descriptions are not diverse. To address this, we introduce ULIP-2, a simple yet effective tri-modal pretraining framework that leverages large multimodal models to automatically generate holistic language descriptions for 3D shapes. It only needs 3D data as input, eliminating the need for any manual 3D annotations, and is therefore scalable to large datasets. ULIP-2 is also equipped with scaled-up backbones for better multimodal representation learning. We conduct experiments on two large-scale 3D datasets, Objaverse and ShapeNet, and augment them with tri-modal datasets of 3D point clouds, images, and language for training ULIP-2. Experiments show that ULIP-2 demonstrates substantial benefits in three downstream tasks: zero-shot 3D classification, standard 3D classification with fine-tuning, and 3D captioning (3D-to-language generation). It achieves a new SOTA of 50.6% (top-1) on Objaverse-LVIS and 84.7% (top-1) on ModelNet40 in zero-shot classification. In the ScanObjectNN benchmark for standard fine-tuning, ULIP-2 reaches an overall accuracy of 91.5% with a compact model of only 1.4 million parameters. ULIP-2 sheds light on a new paradigm for scalable multimodal 3D representation learning without human annotations and shows significant improvements over existing baselines. The code and datasets are released at https://github.com/salesforce/ULIP.

While large language models based on the transformer architecture have demonstrated remarkable in-context learning (ICL) capabilities, understandings of such capabilities are still in an early stage, where existing theory and mechanistic understanding focus mostly on simple scenarios such as learning simple function classes. This paper takes initial steps on understanding ICL in more complex scenarios, by studying learning with \emph{representations}. Concretely, we construct synthetic in-context learning problems with a compositional structure, where the label depends on the input through a possibly complex but \emph{fixed} representation function, composed with a linear function that \emph{differs} in each instance. By construction, the optimal ICL algorithm first transforms the inputs by the representation function, and then performs linear ICL on top of the transformed dataset. We show theoretically the existence of transformers that approximately implement such algorithms with mild depth and size. Empirically, we find trained transformers consistently achieve near-optimal ICL performance in this setting, and exhibit the desired dissection where lower layers transforms the dataset and upper layers perform linear ICL. Through extensive probing and a new pasting experiment, we further reveal several mechanisms within the trained transformers, such as concrete copying behaviors on both the inputs and the representations, linear ICL capability of the upper layers alone, and a post-ICL representation selection mechanism in a harder mixture setting. These observed mechanisms align well with our theory and may shed light on how transformers perform ICL in more realistic scenarios.

Recent months have seen the emergence of a powerful new trend in which large language models (LLMs) are augmented to become autonomous language agents capable of performing objective oriented multi-step tasks on their own, rather than merely responding to queries from human users. Most existing language agents, however, are not optimized using environment-specific rewards. Although some agents enable iterative refinement through verbal feedback, they do not reason and plan in ways that are compatible with gradient-based learning from rewards. This paper introduces a principled framework for reinforcing large language agents by learning a retrospective model, which automatically tunes the language agent prompts from environment feedback through policy gradient. Specifically, our proposed agent architecture learns from rewards across multiple environments and tasks, for fine-tuning a pre-trained language model which refines the language agent prompt by summarizing the root cause of prior failed attempts and proposing action plans. Experimental results on various tasks demonstrate that the language agents improve over time and that our approach considerably outperforms baselines that do not properly leverage gradients from the environment.

Deep learning for time series forecasting has traditionally operated within a one-model-per-dataset framework, limiting its potential to leverage the game-changing impact of large pre-trained models. The concept of universal forecasting, emerging from pre-training on a vast collection of time series datasets, envisions a single Large Time Series Model capable of addressing diverse downstream forecasting tasks. However, constructing such a model poses unique challenges specific to time series data: (i) cross-frequency learning, (ii) accommodating an arbitrary number of variates for multivariate time series, and (iii) addressing the varying distributional properties inherent in large-scale data. To address these challenges, we present novel enhancements to the conventional time series Transformer architecture, resulting in our proposed Masked Encoder-based Universal Time Series Forecasting Transformer (Moirai). Trained on our newly introduced Large-scale Open Time Series Archive (LOTSA) featuring over 27B observations across nine domains, Moirai achieves competitive or superior performance as a zero-shot forecaster when compared to full-shot models. Code, data, and model weights can be found at https://github.com/SalesforceAIResearch/uni2ts.

When deploying modern machine learning-enabled robotic systems in high-stakes applications, detecting distribution shift is critical. However, most existing methods for detecting distribution shift are not well-suited to robotics settings, where data often arrives in a streaming fashion and may be very high-dimensional. In this work, we present an online method for detecting distribution shift with guarantees on the false positive rate — i.e., when there is no distribution shift, our system is very unlikely (with probability < ε) to falsely issue an alert; any alerts that are issued should therefore be heeded. Our method is specifically designed for efficient detection even with high dimensional data, and it empirically achieves up to 11x faster detection on realistic robotics settings compared to prior work while maintaining a low false negative rate in practice (whenever there is a distribution shift in our experiments, our method indeed emits an alert). We demonstrate our approach in both simulation and hardware for a visual servoing task, and show that our method indeed issues an alert before a failure occurs.

Multimodal interleaved datasets featuring free-form interleaved sequences of images and text are crucial for training frontier large multimodal models (LMMs). Despite the rapid progression of open-source LMMs, there remains a pronounced scarcity of large-scale, open-source multimodal interleaved datasets.In response, we introduce MINT-1T, the most extensive and diverse open-source Multimodal INTerleaved dataset to date. MINT-1T comprises of one trillion text tokens and 3.4 billion images, a 10x scale-up from existing open-source datasets. Additionally, we include previously untapped sources such as PDFs and ArXiv papers. As scaling multimodal interleaved datasets requires substantial engineering effort, sharing the data curation process and releasing the dataset greatly benefits the community. Our experiments show that LMMs trained on MINT-1T rival the performance of models trained on the previous leading dataset, OBELICS. We release our data at https://github.com/mlfoundations/MINT-1T.

Large language models (LLMs) for code are typically trained to align with natural language instructions to closely follow their intentions and requirements. However, in many practical scenarios, it becomes increasingly challenging for these models to navigate the intricate boundary between helpfulness and safety, especially against highly complex yet potentially malicious instructions. In this work, we introduce INDICT: a new framework that empowers LLMs with Internal Dialogues of Critiques for both safety and helpfulness guidance. The internal dialogue is a dual cooperative system between a safety-driven critic and a helpfulness-driven critic. Each critic provides analysis against the given task and corresponding generated response, equipped with external knowledge queried through relevant code snippets and tools like web search and code interpreter. We engage the dual critic system in both code generation stage as well as code execution stage, providing preemptive and post-hoc guidance respectively to LLMs. We evaluated INDICT on 8 diverse tasks across 8 programming languages from 5 benchmarks, using LLMs from 7B to 70B parameters. We observed that our approach can provide an advanced level of critiques of both safety and helpfulness analysis, significantly improving the quality of output codes (+10% absolute improvements in all models).

The advancement of function-calling agent models requires diverse, reliable, and high-quality datasets. This paper presents APIGen, an automated data generation pipeline designed to synthesize high-quality datasets for function-calling applications. We leverage APIGen and collect 3,673 executable APIs across 21 different categories to generate diverse function-calling datasets in a scalable and structured manner. Each data in our dataset is verified through three hierarchical stages: format checking, actual function executions, and semantic verification, improving its reliability and correctness. We demonstrate that models trained with our curated datasets, even with only 7B parameters, can achieve state-of-the-art performance on the Berkeley Function-Calling Benchmark, outperforming multiple GPT-4 models. Moreover, our 1B model achieves exceptional performance, surpassing GPT-3.5-Turbo and Claude-3 Haiku. We release a dataset containing 60,000 high-quality entries, aiming to advance the field of function-calling agent domains. The dataset and models are available on the project homepage \url{https://apigen-pipeline.github.io/}.

Autonomous agents that accomplish complex computer tasks with minimal human interventions have the potential to transform human-computer interaction, significantly enhancing accessibility and productivity. However, existing benchmarks either lack an interactive environment or are limited to environments specific to certain applications or domains, failing to reflect the diverse and complex nature of real-world computer use, thereby limiting the scope of tasks and agent scalability. To address this issue, we introduce OSWorld, the first-of-its-kind scalable, real computer environment for multimodal agents, supporting task setup, execution-based evaluation, and interactive learning across various operating systems such as Ubuntu, Windows, and macOS. OSWorld can serve as a unified, integrated computer environment for assessing open-ended computer tasks that involve arbitrary applications. Building upon OSWorld, we create a benchmark of 369 computer tasks involving real web and desktop apps in open domains, OS file I/O, and workflows spanning multiple applications. Each task example is derived from real-world computer use cases and includes a detailed initial state setup configuration and a custom execution-based evaluation script for reliable, reproducible evaluation. Extensive evaluation of state-of-the-art LLM/VLM-based agents on OSWorld reveals significant deficiencies in their ability to serve as computer assistants. While humans can accomplish over 72.36% of the tasks, the best model achieves only 12.24% success, primarily struggling with GUI grounding and operational knowledge. Comprehensive analysis using OSWorld provides valuable insights for developing multimodal generalist agents that were not possible with previous benchmarks. Our code, environment, baseline models, and data are publicly available at this https URL.

Modern natural language generation paradigms require a decoding strategy to obtain quality sequences out of the model.Beam search yields high-quality but low diversity outputs; stochastic approaches suffer from high variance and sometimes low quality.In this work, we propose a deterministic search algorithm balancing both quality and diversity.We first investigate the vanilla best-first search (BFS) algorithm and then propose the best-k search algorithm.Inspired by BFS, we greedily expand the top k nodes, instead of the first node, to boost efficiency and diversity.Upweighting recently discovered nodes accompanied by heap pruning ensures the completeness of the search procedure.Experiments on four NLG tasks show that best-k search yields more diverse and natural outputs compared to strong baselines, while our approach maintains high text quality.The proposed algorithm is parameter-free, lightweight, efficient, and easy-to-use.1

The recognition capabilities of current state-of-the-art 3D models are limited by datasets with a small number of annotated data and a pre-defined set of categories. In its 2D counterpart, recent advances have shown that similar problems can be significantly alleviated by employing knowledge from other modalities, such as language. Inspired by this, leveraging multimodal information for 3D modality could be promising to improve 3D understanding under the restricted data regime, but this line of research is not well studied. Therefore, we introduce ULIP to learn a unified representation of image, text, and 3D point cloud by pre-training with object triplets from the three modalities. To overcome the shortage of training triplets, ULIP leverages a pre-trained vision-language model that has already learned a common visual and textual space by training with massive image-text pairs. Then, ULIP learns a 3D representation space aligned with the common image-text space, using a small number of automatically synthesized triplets. ULIP is agnostic to 3D backbone networks and can easily be integrated into any 3D architecture. Experiments show that ULIP effectively improves the performance of multiple recent 3D backbones by simply pre-training them on ShapeNet55 using our framework, achieving state-of-the-art performance in both standard 3D classification and zero-shot 3D classification on ModelNet40 and ScanObjectNN. ULIP also improves the performance of PointMLP by around 3% in 3D classification on ScanObjectNN, and outperforms PointCLIP by 28.8% on top-1 accuracy for zero-shot 3D classification on ModelNet40. Our code and pre-trained models will be released.

Our goal is to learn a video representation that is useful for downstream procedure understanding tasks in instructional videos. Due to the small amount of available annotations, a key challenge in procedure understanding is to be able to extract from unlabeled videos the procedural knowledge such as the identity of the task (e.g., ‘make latte’), its steps (e.g., ‘pour milk’), or the potential next steps given partial progress in its execution. Our main insight is that instructional videos depict sequences of steps that repeat between instances of the same or different tasks, and that this structure can be well represented by a Procedural Knowledge Graph (PKG), where nodes are discrete steps and edges connect steps that occur sequentially in the instructional activities. This graph can then be used to generate pseudo labels to train a video representation that encodes the procedural knowledge in a more accessible form to generalize to multiple procedure understanding tasks. We build a PKG by combining information from a text-based procedural knowledge database and an unlabeled instructional video corpus and then use it to generate training pseudo labels with four novel pre-training objectives. We call this PKG-based pre-training procedure and the resulting model Paprika, Procedure-Aware PRe-training for Instructional Knowledge Acquisition. We evaluate Paprika on COIN and CrossTask for procedure understanding tasks such as task recognition, step recognition, and step forecasting. Paprika yields a video representation that improves over the state of the art: up to 11.23% gains in accuracy in 12 evaluation settings. Implementation is available at https://github.com/salesforce/paprika.

The cost of vision-and-language pre-training has become increasingly prohibitive due to end-to-end training of large-scale models. This paper proposes BLIP-2, a generic and efficient pre-training strategy that bootstraps vision-language pre-training from off-the-shelf frozen pre-trained image encoders and frozen large language models. BLIP-2 bridges the modality gap with a lightweight Querying Transformer, which is pre-trained in two stages. The first stage bootstraps vision-language representation learning from a frozen image encoder. The second stage bootstraps vision-to-language generative learning from a frozen language model. BLIP-2 achieves state-of-the-art performance on various vision-language tasks, despite having significantly fewer trainable parameters than existing methods. For example, our model outperforms Flamingo80B by 8.7% on zero-shot VQAv2 with 54x fewer trainable parameters. We also demonstrate the model’s emerging capabilities of zero-shot image-to-text generation that can follow natural language instructions.

Embodied AI agents that search for objects in large environments such as households often need to make efficient decisions by predicting object locations based on partial information. We pose this as a new type of link prediction problem: link prediction on partially observable dynamic graphs Our graph is a representation of a scene in which rooms and objects are nodes, and their relationships are encoded in the edges; only parts of the changing graph are known to the agent at each timestep. This partial observability poses a challenge to existing link prediction approaches, which we address. We propose a novel state representation – Scene Graph Memory (SGM) – with captures the agent’s accumulated set of observations, as well as a neural net architecture called a Node Edge Predictor (NEP) that extracts information from the SGM to search efficiently. We evaluate our method in the Dynamic House Simulator, a new benchmark that creates diverse dynamic graphs following the semantic patterns typically seen at homes, and show that NEP can be trained to predict the locations of objects in a variety of environments with diverse object movement dynamics, outperforming baselines both in terms of new scene adaptability and overall accuracy. The codebase and more can be found www.scenegraphmemory.com.

Developing embodied agents in simulation has been a key research topic in recent years. Exciting new tasks, algorithms, and benchmarks have been developed in various simulators. However, most of them assume deaf agents in silent environments, while we humans perceive the world with multiple senses. We introduce Sonicverse, a multisensory simulation platform with integrated audio-visual simulation for training household agents that can both see and hear. Sonicverse models realistic continuous audio rendering in 3D environments in real-time. Together with a new audio-visual VR interface that allows humans to interact with agents with audio, Sonicverse enables a series of embodied AI tasks that need audio-visual perception. For semantic audio-visual navigation in particular, we also propose a new multi-task learning model that achieves state-of-the-art performance. In addition, we demonstrate Sonicverse's realism via sim-to-real transfer, which has not been achieved by other simulators: an agent trained in Sonicverse can successfully perform audio-visual navigation in real-world environments. Sonicverse is available at: https://github.com/StanfordVL/Sonicverse.

Achieving machine autonomy and human control often represent divergent objectives in the design of interactive AI systems. Visual generative foundation models such as Stable Diffusion show promise in navigating these goals, especially when prompted with arbitrary languages. However, they often fall short in generating images with spatial, structural, or geometric controls. The integration of such controls, which can accommodate various visual conditions in a single unified model, remains an unaddressed challenge. In response, we introduce UniControl, a new generative foundation model that consolidates a wide array of controllable condition-to-image (C2I) tasks within a singular framework, while still allowing for arbitrary language prompts. UniControl enables pixel-level-precise image generation, where visual conditions primarily influence the generated structures and language prompts guide the style and context. To equip UniControl with the capacity to handle diverse visual conditions, we augment pretrained text-to-image diffusion models and introduce a task-aware HyperNet to modulate the diffusion models, enabling the adaptation to different C2I tasks simultaneously. Trained on nine unique C2I tasks, UniControl demonstrates impressive zero-shot generation abilities with unseen visual conditions. Experimental results show that UniControl often surpasses the performance of single-task-controlled methods of comparable model sizes. This control versatility positions UniControl as a significant advancement in the realm of controllable visual generation.

Jianguo Zhang, Stephen Roller, Kun Qian, Zhiwei Liu, Rui Meng, Shelby Heinecke, Huan Wang, Silvio Savarese, Caiming Xiong. Proceedings of the 24th Meeting of the Special Interest Group on Discourse and Dialogue. 2023.

We introduce Merlion, an open-source machine learning library for time series. It features a unified interface for many commonly used models and datasets for forecasting and anomaly detection on both univariate and multivariate time series, along with standard pre/post-processing layers. It has several modules to improve ease-of-use, including a no-code visual dashboard, anomaly score calibration to improve interpetability, AutoML for hyperparameter tuning and model selection, and model ensembling. Merlion also provides an evaluation framework that simulates the live deployment of a model in production, and a distributed computing backend to run time series models at industrial scale. This library aims to provide engineers and researchers a one-stop solution to rapidly develop models for their specific time series needs and benchmark them across multiple datasets.

We present JRDB, a novel egocentric dataset collected from our social mobile manipulator JackRabbot. The dataset includes 64 minutes of annotated multimodal sensor data including stereo cylindrical 360${}^\circ$RGB video at 15 fps, 3D point clouds from two 16 planar rays Velodyne LiDARs, line 3D point clouds from two Sick Lidars, audio signal, RGB-D video at 30 fps, 360${}^\circ$spherical image from a fisheye camera and encoder values from the robot's wheels. Our dataset incorporates data from traditionally underrepresented scenes such as indoor environments and pedestrian areas, all from the ego-perspective of the robot, both stationary and navigating. The dataset has been annotated with over 2.4 million bounding boxes spread over five individual cameras and 1.8 million associated 3D cuboids around all people in the scenes totaling over 3500 time consistent trajectories. Together with our dataset and the annotations, we launch a benchmark and metrics for 2D and 3D person detection and tracking. With this dataset, which we plan on extending with further types of annotation in the future, we hope to provide a new source of data and a test-bench for research in the areas of egocentric robot vision, autonomous navigation, and all perceptual tasks around social robotics in human environments.

This article examines performance evaluation criteria for basic vision tasks involving sets of objects namely, object detection, instance-level segmentation and multi-object tracking. The rankings of algorithms by a criterion can fluctuate with different choices of parameters, e.g. Intersection over Union (IoU) threshold, making their evaluations unreliable. More importantly, there is no means to verify whether we can trust the evaluations of a criterion. This work suggests a notion of trustworthiness for performance criteria, which requires (i) robustness to parameters for reliability, (ii) contextual meaningfulness in sanity tests, and (iii) consistency with mathematical requirements such as the metric properties. We observe that these requirements were overlooked by many widely-used criteria, and explore alternative criteria using metrics for sets of shapes. We also assess all these criteria based on the suggested requirements for trustworthiness.

The availability of large-scale video action understanding datasets has facilitated advances in the interpretation of visual scenes containing people. However, learning to recognise human actions and their social interactions in an unconstrained real-world environment comprising numerous people, with potentially highly unbalanced and longtailed distributed action labels from a stream of sensory data captured from a mobile robot platform remains a significant challenge, not least owing to the lack of a reflective large-scale dataset. In this paper, we introduce JRDB-Act, as an extension of the existing JRDB, which is captured by a social mobile manipulator and reflects a real distribution of human daily-life actions in a university campus environment. JRDB-Act has been densely annotated with atomic actions, comprises over 2.8M action labels, constituting a large-scale spatio-temporal action detection dataset. Each human bounding box is labeled with one pose-based action label and multiple (optional) interaction-based action labels. Moreover JRDB-Act provides social group annotation, conducive to the task of grouping individuals based on their interactions in the scene to infer their social activities (common activities in each social group). Each annotated label in JRDB-Act is tagged with the annotators' confidence level which contributes to the development of reliable evaluation strategies. In order to demonstrate how one can effectively utilise such annotations, we develop an end-to-end trainable pipeline to learn and infer these tasks, i.e. individual action and social group detection. The data and the evaluation code will be publicly available at https://jrdb.erc.monash.edu/

Program synthesis or code generation aims to generate a program that satisfies a problem specification. Recent approaches using large-scale pretrained language models (LMs) have shown promising results, yet they have some critical limitations. In particular, they often follow a standard supervised fine-tuning procedure to train a code generation model from natural language problem descriptions and ground-truth programs only. Such paradigm largely ignores some important but potentially useful signals in the problem specification such as unit tests, which thus results in poor performance when solving complex unseen coding tasks. We propose “CodeRL” to address the limitations, a new framework for program synthesis tasks through pretrained LMs and deep reinforcement learning (RL). Specifically, during training, we treat the code-generating LM as an actor network, and introduce a critic network that is trained to predict the functional correctness of generated programs and provide dense feedback signals to the actor. During inference, we introduce a new generation procedure with a critical sampling strategy that allows a model to automatically regenerate programs based on feedback from example unit tests and critic scores. For the model backbones, we extended the encoder-decoder architecture of CodeT5 with enhanced learning objectives, larger model sizes, and better pretraining data. Our method not only achieves new SOTA results on the challenging APPS benchmark, but also shows strong zero-shot transfer capability with new SOTA results on the simpler MBPP benchmark.

Probabilistic classifiers output confidence scores along with their predictions, and these confidence scores should be calibrated, i.e., they should reflect the reliability of the prediction. Confidence scores that minimize standard metrics such as the expected calibration error (ECE) accurately measure the reliability \textit{on average} across the entire population. However, it is in general impossible to measure the reliability of an \textit{individual} prediction. In this work, we propose the local calibration error (LCE) to span the gap between average and individual reliability. For each individual prediction, the LCE measures the average reliability of a set of similar predictions, where similarity is quantified by a kernel function on a pretrained feature space and by a binning scheme over predicted model confidences. We show theoretically that the LCE can be estimated sample-efficiently from data, and empirically find that it reveals miscalibration modes that are more fine-grained than the ECE can detect. Our key result is a novel {\textbf{lo}cal \textbf{re}calibration} method \method{}, to improve confidence scores for individual predictions and decrease the LCE. Experimentally, we show that our recalibration method produces more accurate confidence scores, which improves downstream fairness and decision making on classification tasks with both image and tabular data.

Conventional approaches to vision-and-language navigation (VLN) are trained end-to-end but struggle to perform well in freely traversable environments. Inspired by the robotics community, we propose a modular approach to VLN using topological maps. Given a natural language instruction and topological map, our approach leverages attention mechanisms to predict a navigation plan in the map. The plan is then executed with low-level actions (e.g. FORWARD, ROTATE) using a robust controller. Experiments show that our method outperforms previous end-to-end approaches, generates interpretable navigation plans, and exhibits intelligent behaviors such as backtracking.

Joint forecasting of human trajectory and pose dynamics is a fundamental building block of various applications ranging from robotics and autonomous driving to surveillance systems. Predicting body dynamics requires capturing subtle information embedded in the humans’ interactions with each other and with the objects present in the scene. In this paper, we propose a novel TRajectory and POse Dynamics (nicknamed TRiPOD) method based on graph attentional networks to model the human-human and human-object interactions both in the input space and the output space (decoded future output). The model is supplemented by a message passing interface over the graphs to fuse these different levels of interactions efficiently. Furthermore, to incorporate a real-world challenge, we propound to learn an indicator representing whether an estimated body joint is visible/invisible at each frame, e.g. due to occlusion or being outside the sensor field of view. Finally, we introduce a new benchmark for this joint task based on two challenging datasets (PoseTrack and 3DPW) and propose evaluation metrics to measure the effectiveness of predictions in the global space, even when there are invisible cases of joints. Our evaluation shows that TRiPOD outperforms all prior work and state-of-the-art specifically designed for each of the trajectory and pose forecasting tasks.

The process of learning a manipulation task depends strongly on the action space used for exploration: posed in the incorrect action space, solving a task with reinforcement learning can be drastically inefficient. Additionally, similar tasks or instances of the same task family impose latent manifold constraints on the most effective action space: the task family can be best solved with actions in a manifold of the entire action space of the robot. Combining these insights we present LASER, a method to learn latent action spaces for efficient reinforcement learning. LASER factorizes the learning problem into two sub-problems, namely action space learning and policy learning in the new action space. It leverages data from similar manipulation task instances, either from an offline expert or online during policy learning, and learns from these trajectories a mapping from the original to a latent action space. LASER is trained as a variational encoder-decoder model to map raw actions into a disentangled latent action space while maintaining action reconstruction and latent space dynamic consistency. We evaluate LASER on two contact-rich robotic tasks in simulation, and analyze the benefit of policy learning in the generated latent action space. We show improved sample efficiency compared to the original action space from better alignment of the action space to the task space, as we observe with visualizations of the learned action space manifold. Additional details: pair.toronto.edu/laser

Searching for objects in indoor organized environments such as homes or offices is part of our everyday activities. When looking for a desired object, we reason about the rooms and containers the object is likely to be in; the same type of container will have a different probability of containing the target depending on which room it is in. We also combine geometric and semantic information to infer what container is best to search, or what other objects are best to move, if the target object is hidden from view. We use a 3D scene graph representation to capture the hierarchical, semantic, and geometric aspects of this problem. To exploit this representation in a search process, we introduce Hierarchical Mechanical Search (HMS), a method that guides an agent’s actions towards finding a target object specified with a natural language description. HMS is based on a novel neural network architecture that uses neural message passing of vectors with visual, geometric, and linguistic information to allow HMS to process data across layers of the graph while combining semantic and geometric cues. HMS is trained on 1000 3D scene graphs and evaluated on a novel dataset of 500 3D scene graphs with dense placements of semantically related objects in storage locations, and is shown to be significantly better than several baselines at finding objects. It is also close to the oracle policy in terms of the median number of actions required. Additional qualitative results can be found at https://ai.stanford.edu/mech-search/hms

Navigating fluently around pedestrians is a necessary capability for mobile robots deployed in human environments, such as buildings and homes. While research on social navigation has focused mainly on the scalability with the number of pedestrians in open spaces, typical indoor environments present the additional challenge of constrained spaces such as corridors and doorways that limit maneuverability and influence patterns of pedestrian interaction. We present an approach based on reinforcement learning (RL) to learn policies capable of dynamic adaptation to the presence of moving pedestrians while navigating between desired locations in constrained environments. The policy network receives guidance from a motion planner that provides waypoints to follow a globally planned trajectory, whereas RL handles the local interactions. We explore a compositional principle for multi-layout training and find that policies trained in a small set of geometrically simple layouts successfully generalize to more complex unseen layouts that exhibit composition of the structural elements available during training. Going beyond walls-world like domains, we show transfer of the learned policy to unseen 3D reconstructions of two real environments. These results support the applicability of the compositional principle to navigation in real-world buildings and indicate promising usage of multi-agent simulation within reconstructed environments for tasks that involve interaction. https://ai.stanford.edu/∼cdarpino/socialnavconstrained/

Imitation Learning (IL) is a powerful paradigm to teach robots to perform manipulation tasks by allowing them to learn from human demonstrations collected via teleoperation, but has mostly been limited to single-arm manipulation. However, many real-world tasks require multiple arms, such as lifting a heavy object or assembling a desk. Unfortunately, applying IL to multi-arm manipulation tasks has been challenging –asking a human to control more than one robotic arm can impose significant cognitive burden and is often only possible for a maximum of two robot arms. To address these challenges, we present MULTI-ARM ROBOTURK (MART), a multi-user data collection platform that allows multiple remote users to simultaneously teleoperate a set of robotic arms and collect demonstrations for multi-arm tasks. Using MART, we collected demonstrations for five novel two and three-arm tasks from several geographically separated users. From our data we arrived at a critical insight: most multi-arm tasks do not require global coordination throughout its full duration, but only during specific moments. We show that learning from such data consequently presents challenges for centralized agents that directly attempt to model all robot actions simultaneously, and perform a comprehensive study of different policy architectures with varying levels of centralization on our tasks. Finally, we propose and evaluate a base-residual policy framework that allows trained policies to better adapt to the mixed coordination setting common in multi-arm manipulation, and show that a centralized policy augmented with a decentralized residual model outperforms all other models on our set of benchmark tasks. Additional results and videos at https://roboturk.stanford.edu/multiarm

Many Reinforcement Learning (RL) approaches use joint control signals (positions, velocities, torques) as action space for continuous control tasks. We propose to lift the action space to a higher level in the form of subgoals for a motion generator (a combination of motion planner and trajectory executor). We argue that, by lifting the action space and by leveraging sampling-based motion planners, we can efficiently use RL to solve complex, long-horizon tasks that could not be solved with existing RL methods in the original action space. We propose ReLMoGen – a framework that combines a learned policy to predict subgoals and a motion generator to plan and execute the motion needed to reach these subgoals. To validate our method, we apply ReLMoGen to two types of tasks: 1) Interactive Navigation tasks, navigation problems where interactions with the environment are required to reach the destination, and 2) Mobile Manipulation tasks, manipulation tasks that require moving the robot base. These problems are challenging because they are usually long-horizon, hard to explore during training, and comprise alternating phases of navigation and interaction. Our method is benchmarked on a diverse set of seven robotics tasks in photo-realistic simulation environments. In all settings, ReLMoGen outperforms state-of-the-art RL and Hierarchical RL baselines. ReLMoGen also shows outstanding transferability between different motion generators at test time, indicating a great potential to transfer to real robots. For more information, please visit project website: http://svl.stanford.edu/projects/relmogen.

Planning in realistic environments requires searching in large planning spaces. Affordances are a powerful concept to simplify this search, because they model what actions can be successful in a given situation. However, the classical notion of affordance is not suitable for long horizon planning because it only informs the robot about the immediate outcome of actions instead of what actions are best for achieving a long-term goal. In this paper, we introduce a new affordance representation that enables the robot to reason about the longterm effects of actions through modeling what actions are afforded in the future. Based on the new representation, we develop a learning-to-plan method, Deep Affordance Foresight (DAF), that learns partial environment models of affordances of parameterized motor skills through trial-and-error. We evaluate DAF on two challenging manipulation domains and show that it can effectively learn to carry out multi-step tasks, share learned affordance representations among different tasks, and learn to plan with high-dimensional image inputs.

Humans can robustly follow a visual trajectory defined by a sequence of images (i.e. a video) regardless of substantial changes in the environment or the presence of obstacles. We aim at endowing similar visual navigation capabilities to mobile robots solely equipped with a RGB fisheye camera. We propose a novel probabilistic visual navigation system that learns to follow a sequence of images with bidirectional visual predictions conditioned on possible navigation velocities. By predicting bidirectionally (from start towards goal and vice versa) our method extends its predictive horizon enabling the robot to go around unseen large obstacles that are not visible in the video trajectory. Learning how to react to obstacles and potential risks in the visual field is achieved by imitating human teleoperators. Since the human teleoperation commands are diverse, we propose a probabilistic representation of trajectories that we can sample to find the safest path. We evaluate our navigation system quantitatively and qualitatively in multiple simulated and real environments and compare to state-of-the-art baselines. Our approach outperforms the most recent visual navigation methods with a large margin with regard to goal arrival rate, subgoal coverage rate, and success weighted by path length (SPL). Our method also generalizes to new robot embodiments never used during training.

We present iGibson 1.0, a novel simulation environment to develop robotic solutions for interactive tasks in large-scale realistic scenes. Our environment contains 15 fully interactive home-sized scenes with 108 rooms populated with rigid and articulated objects. The scenes are replicas of real-world homes, with distribution and the layout of objects aligned to those of the real world. iGibson 1.0 integrates several key features to facilitate the study of interactive tasks: i) generation of high-quality virtual sensor signals (RGB, depth, segmentation, LiDAR, flow and so on), ii) domain randomization to change the materials of the objects (both visual and physical) and/or their shapes, iii) integrated sampling-based motion planners to generate collision-free trajectories for robot bases and arms, and iv) intuitive human-iGibson interface that enables efficient collection of human demonstrations. Through experiments, we show that the full interactivity of the scenes enables agents to learn useful visual representations that accelerate the training of downstream manipulation tasks. We also show that iGibson features enable the generalization of navigation agents, and that the human-iGibson interface and integrated motion planners facilitate efficient imitation learning of human demonstrated (mobile) manipulation behaviors. iGibson 1.0 is open-source, equipped with comprehensive examples and documentation. For more information, visit our project website: http://svl.stanford.edu/igibson/.

Imitation Learning (IL) is an effective framework to learn visuomotor skills from offline demonstration data. However, IL methods often fail to generalize to new scene configurations not covered by training data. On the other hand, humans can manipulate objects in varying conditions. Key to such capability is hand-eye coordination, a cognitive ability that enables humans to adaptively direct their movements at task-relevant objects and be invariant to the objects’ absolute spatial location. In this work, we present a learnable action space, Hand-eye Action Networks (HAN) that learns coordinated hand-eye movements from human teleoperated demonstrations. Through a set of challenging multi-stage manipulation tasks, we show that a visuomotor policy equipped with HAN is able to inherit the key spatial invariance property of handeye coordination and achieve generalization to new scene configurations. Additional materials available at https://sites.google.com/stanford.edu/han

Learned dynamics models combined with both planning and policy learning algorithms have shown promise in enabling artificial agents to learn to perform many diverse tasks with limited supervision. However, one of the fundamental challenges in using a learned forward dynamics model is the mismatch between the objective of the learned model (future state reconstruction), and that of the downstream planner or policy (completing a specified task). This issue is exacerbated by vision-based control tasks in diverse real-world environments, where the complexity of the real world dwarfs model capacity. In this paper, we propose to direct prediction towards task relevant information, enabling the model to be aware of the current task and encouraging it to only model relevant quantities of the state space, resulting in a learning objective that more closely matches the downstream task. Further, we do so in an entirely self-supervised manner, without the need for a reward function or image labels. We find that our method more effectively models the relevant parts of the scene conditioned on the goal, and as a result outperforms standard task-agnostic dynamics models and model-free reinforcement learning.

Many computer vision applications require solving multiple tasks in real-time. A neural network can be trained to solve multiple tasks simultaneously using multi-task learning. This can save computation at inference time as only a single network needs to be evaluated. Unfortunately, this often leads to inferior overall performance as task objectives can compete, which consequently poses the question: which tasks should and should not be learned together in one network when employing multi-task learning? We study task cooperation and competition in several different learning settings and propose a framework for assigning tasks to a few neural networks such that cooperating tasks are computed by the same neural network, while competing tasks are computed by different networks. Our framework offers a time-accuracy trade-off and can produce better accuracy using less inference time than not only a single large multi-task neural network but also many single-task networks.

Learning from offline task demonstrations is a problem of great interest in robotics. For simple short-horizon manipulation tasks with modest variation in task instances, offline learning from a small set of demonstrations can produce controllers that successfully solve the task. However, leveraging a fixed batch of data can be problematic for larger datasets and longer-horizon tasks with greater variations. The data can exhibit substantial diversity and consist of suboptimal solution approaches. In this paper, we propose Implicit Reinforcement without Interaction at Scale (IRIS), a novel framework for learning from large-scale demonstration datasets. IRIS factorizes the control problem into a goal-conditioned low-level controller that imitates short demonstration sequences and a high-level goal selection mechanism that sets goals for the low-level and selectively combines parts of suboptimal solutions leading to more successful task completions. We evaluate IRIS across three datasets, including the RoboTurk Cans dataset collected by humans via crowdsourcing, and show that performant policies can be learned from purely offline learning. Additional results at https://sites.google.com/stanford.edu/iris/.

We aim to develop an algorithm for robots to manipulate novel objects as tools for completing different task goals. An efficient and informative representation would facilitate the effectiveness and generalization of such algorithms. For this purpose, we present KETO, a framework of learning keypoint representations of tool-based manipulation. For each task, a set of task-specific keypoints is jointly predicted from 3D point clouds of the tool object by a deep neural network. These keypoints offer a concise and informative description of the object to determine grasps and subsequent manipulation actions. The model is learned from self-supervised robot interactions in the task environment without the need for explicit human annotations. We evaluate our framework in three manipulation tasks with tool use. Our model consistently outperforms state-of-the-art methods in terms of task success rates. Qualitative results of keypoint prediction and tool generation are shown to visualize the learned representations.

We present 6-PACK, a deep learning approach to category-level 6D object pose tracking on RGB-D data. Our method tracks in real time novel object instances of known object categories such as bowls, laptops, and mugs. 6-PACK learns to compactly represent an object by a handful of 3D keypoints, based on which the interframe motion of an object instance can be estimated through keypoint matching. These keypoints are learned end-to-end without manual supervision in order to be most effective for tracking. Our experiments show that our method substantially outperforms existing methods on the NOCS category-level 6D pose estimation benchmark and supports a physical robot to perform simple vision-based closed-loop manipulation tasks. Our code and video are available at https://sites.google.com/view/6packtracking.

We propose a method to facilitate robot navigation relative to sketched maps of human environments. Our main contribution centers around using thin plate splines for registering the robot's LIDAR observation with the hand-drawn maps. Thin plate splines are particularly effective for this task because they are able to handle many of the nonrigid deformations commonly seen in sketches of maps, which render traditional rigid transformations inappropriate. Our proposed approach uses a convolutional neural network to efficiently predict the control points which define the spline transform, from which we then compute the pose of the robot on the hand drawn map for navigation purposes. Our systematic evaluations in simulation using a synthetic dataset and real, hand-drawn sketches show that the proposed spline-based registration approach outperforms baseline methods.

When searching for objects in cluttered environments, it is often necessary to perform complex interactions in order to move occluding objects out of the way and fully reveal the object of interest and make it graspable. Due to the complexity of the physics involved and the lack of accurate models of the clutter, planning and controlling precise predefined interactions with accurate outcome is extremely hard, when not impossible. In problems where accurate (forward) models are lacking, Deep Reinforcement Learning (RL) has shown to be a viable solution to map observations (e.g. images) to good interactions in the form of close-loop visuomotor policies. However, Deep RL is sample inefficient and fails when applied directly to the problem of unoccluding objects based on images. In this work we present a novel Deep RL procedure that combines i) teacher-aided exploration, ii) a critic with privileged information, and iii) mid-level representations, resulting in sample efficient and effective learning for the problem of uncovering a target object occluded by a heap of unknown objects. Our experiments show that our approach trains faster and converges to more efficient uncovering solutions than baselines and ablations, and that our uncovering policies lead to an average improvement in the graspability of the target object, facilitating downstream retrieval applications.

Leveraging multimodal information with recursive Bayesian filters improves performance and robustness of state estimation, as recursive filters can combine different modalities according to their uncertainties. Prior work has studied how to optimally fuse different sensor modalities with analytical state estimation algorithms. However, deriving the dynamics and measurement models along with their noise profile can be difficult or lead to intractable models. Differentiable filters provide a way to learn these models end-to-end while retaining the algorithmic structure of recursive filters. This can be especially helpful when working with sensor modalities that are high dimensional and have very different characteristics. In contact-rich manipulation, we want to combine visual sensing (which gives us global information) with tactile sensing (which gives us local information). In this paper, we study new differentiable filtering architectures to fuse heterogeneous sensor information. As case studies, we evaluate three tasks: two in planar pushing (simulated and real) and one in manipulating a kinematically constrained door (simulated). In extensive evaluations, we find that differentiable filters that leverage crossmodal sensor information reach comparable accuracies to unstructured LSTM models, while presenting interpretability benefits that may be important for safety-critical systems. We also release an open-source library for creating and training differentiable Bayesian filters in PyTorch, which can be found on our project website: https://sites.google.com/view/multimodalfilter.

Robots navigating autonomously need to perceive and track the motion of objects and other agents in its surroundings. This information enables planning and executing robust and safe trajectories. To facilitate these processes, the motion should be perceived in 3D Cartesian space. However, most recent multi-object tracking (MOT) research has focused on tracking people and moving objects in 2D RGB video sequences. In this work we present JRMOT, a novel 3D MOT system that integrates information from RGB images and 3D point clouds to achieve real-time, state-of-the-art tracking performance. Our system is built with recent neural networks for re-identification, 2D and 3D detection and track description, combined into a joint probabilistic data-association framework within a multi-modal recursive Kalman architecture. As part of our work, we release the JRDB dataset, a novel large scale 2D+3D dataset and benchmark, annotated with over 2 million boxes and 3500 time consistent 2D+3D trajectories across 54 indoor and outdoor scenes. JRDB contains over 60 minutes of data including 360ocylindrical RGB video and 3D pointclouds in social settings that we use to develop, train and evaluate JRMOT. The presented 3D MOT system demonstrates state-of-the-art performance against competing methods on the popular 2D tracking KITTI benchmark and serves as first 3D tracking solution for our benchmark. Real-robot tests on our social robot JackRabbot indicate that the system is capable of tracking multiple pedestrians fast and reliably. We provide the ROS code of our tracker at https://sites.google.com/view/jrmot.

We propose a data-driven approach to detect conversational groups by identifying spatial arrangements typical of these focused social encounters. Our approach uses a novel Deep Affinity Network (DANTE) to predict the likelihood that two individuals in a scene are part of the same conversational group, considering their social context. The predicted pair-wise affinities are then used in a graph clustering framework to identify both small (e.g., dyads) and large groups. The results from our evaluation on multiple, established benchmarks suggest that combining powerful deep learning methods with classical clustering techniques can improve the detection of conversational groups in comparison to prior approaches. Finally, we demonstrate the practicality of our approach in a human-robot interaction scenario. Our efforts show that our work advances group detection not only in theory, but also in practice.

Contact-rich manipulation tasks in unstructured environments often require both haptic and visual feedback. It is nontrivial to manually design a robot controller that combines these modalities, which have very different characteristics. While deep reinforcement learning has shown success in learning control policies for high-dimensional inputs, these algorithms are generally intractable to train directly on real robots due to sample complexity. In this article, we use self-supervision to learn a compact and multimodal representation of our sensory inputs, which can then be used to improve the sample efficiency of our policy learning. Evaluating our method on a peg insertion task, we show that it generalizes over varying geometries, configurations, and clearances, while being robust to external perturbations. We also systematically study different self-supervised learning objectives and representation learning architectures. Results are presented in simulation and on a physical robot.

In many robotics and VR/AR applications, 3D-videos are readily-available input sources (a sequence of depth images, or LIDAR scans). However, in many cases, the 3D-videos are processed frame-by-frame either through 2D convnets or 3D perception algorithms. In this work, we propose 4-dimensional convolutional neural networks for spatio-temporal perception that can directly process such 3D-videos using high-dimensional convolutions. For this, we adopt sparse tensors and propose generalized sparse convolutions that encompass all discrete convolutions. To implement the generalized sparse convolution, we create an open-source auto-differentiation library for sparse tensors that provides extensive functions for high-dimensional convolutional neural networks. We create 4D spatio-temporal convolutional neural networks using the library and validate them on various 3D semantic segmentation benchmarks and proposed 4D datasets for 3D-video perception. To overcome challenges in 4D space, we propose the hybrid kernel, a special case of the generalized sparse convolution, and trilateral-stationary conditional random fields that enforce spatio-temporal consistency in the 7D space-time-chroma space. Experimentally, we show that a convolutional neural network with only generalized 3D sparse convolutions can outperform 2D or 2D-3D hybrid methods by a large margin. Also, we show that on 3D-videos, 4D spatio-temporal convolutional neural networks are robust to noise and outperform the 3D convolutional neural network.

Many robotic applications require the agent to perform long-horizon tasks in partially observable environments. In such applications, decision making at any step can depend on observations received far in the past. Hence, being able to properly memorize and utilize the long-term history is crucial. In this work, we propose a novel memory-based policy, named Scene Memory Transformer (SMT). The proposed policy embeds and adds each observation to a memory and uses the attention mechanism to exploit spatio-temporal dependencies. This model is generic and can be efficiently trained with reinforcement learning over long episodes. On a range of visual navigation tasks, SMT demonstrates superior performance to existing reactive and memory-based policies by a margin.

Our goal is to generate a policy to complete an unseen task given just a single video demonstration of the task in a given domain. We hypothesize that to successfully generalize to unseen complex tasks from a single video demonstration, it is necessary to explicitly incorporate the compositional structure of the tasks into the model. To this end, we propose Neural Task Graph (NTG) Networks, which use conjugate task graph as the intermediate representation to modularize both the video demonstration and the derived policy. We empirically show NTG achieves inter-task generalization on two complex tasks: Block Stacking in BulletPhysics and Object Collection in AI2-THOR. NTG improves data efficiency with visual input as well as achieve strong generalization without the need for dense hierarchical supervision. We further show that similar performance trends hold when applied to real-world data. We show that NTG can effectively predict task structure on the JIGSAWS surgical dataset and generalize to unseen tasks.

Intersection over Union (IoU) is the most popular evaluation metric used in the object detection benchmarks. However, there is a gap between optimizing the commonly used distance losses for regressing the parameters of a bounding box and maximizing this metric value. The optimal objective for a metric is the metric itself. In the case of axis-aligned 2D bounding boxes, it can be shown that IoU can be directly used as a regression loss. However, IoU has a plateau making it infeasible to optimize in the case of non-overlapping bounding boxes. In this paper, we address the this weakness by introducing a generalized version of IoU as both a new loss and a new metric. By incorporating this generalized IoU (GIoU) as a loss into the state-of-the art object detection frameworks, we show a consistent improvement on their performance using both the standard, IoU based, and new, GIoU based, performance measures on popular object detection benchmarks such as PASCAL VOC and MS COCO.

This paper addresses the problem of path prediction for multiple interacting agents in a scene, which is a crucial step for many autonomous platforms such as self-driving cars and social robots. We present SoPhie; an interpretable framework based on Generative Adversarial Network (GAN), which leverages two sources of information, the path history of all the agents in a scene, and the scene context information, using images of the scene. To predict a future path for an agent, both physical and social information must be leveraged. Previous work has not been successful to jointly model physical and social interactions. Our approach blends a social attention mechanism with physical attention that helps the model to learn where to look in a large scene and extract the most salient parts of the image relevant to the path. Whereas, the social attention component aggregates information across the different agent interactions and extracts the most important trajectory information from the surrounding neighbors. SoPhie also takes advantage of GAN to generates more realistic samples and to capture the uncertain nature of the future paths by modeling its distribution. All these mechanisms enable our approach to predict socially and physically plausible paths for the agents and to achieve state-of-the-art performance on several different trajectory forecasting benchmarks.

3D point cloud generation is of great use for 3D scene modeling and understanding. Real-world 3D object point clouds can be properly described by a collection of low-level and high-level structures such as surfaces, geometric primitives, semantic parts,etc. In fact, there exist many different representations of a 3D object point cloud as a set of point groups. Existing frameworks for point cloud genera-ion either do not consider structure in their proposed solutions, or assume and enforce a specific structure/topology,e.g. a collection of manifolds or surfaces, for the generated point cloud of a 3D object. In this work, we pro-pose a novel decoder that generates a structured point cloud without assuming any specific structure or topology on the underlying point set. Our decoder is softly constrained to generate a point cloud following a hierarchical rooted tree structure. We show that given enough capacity and allowing for redundancies, the proposed decoder is very flexible and able to learn any arbitrary grouping of points including any topology on the point set. We evaluate our decoder on the task of point cloud generation for 3D point cloud shape completion. Combined with encoders from existing frameworks, we show that our proposed decoder significantly outperforms state-of-the-art 3D point cloud completion methods on the Shapenet dataset.

A key technical challenge in performing 6D object pose estimation from RGB-D image is to fully leverage the two complementary data sources. Prior works either extract information from the RGB image and depth separately or use costly post-processing steps, limiting their performances in highly cluttered scenes and real-time applications. In this work, we present DenseFusion, a generic framework for estimating 6D pose of a set of known objects from RGB-D images. DenseFusion is a heterogeneous architecture that processes the two data sources individually and uses a novel dense fusion network to extract pixel-wise dense feature embedding, from which the pose is estimated. Furthermore, we integrate an end-to-end iterative pose refinement procedure that further improves the pose estimation while achieving near real-time inference. Our experiments show that our method outperforms state-of-the-art approaches in two datasets, YCB-Video and LineMOD. We also deploy our proposed method to a real robot to grasp and manipulate objects based on the estimated pose.

A comprehensive semantic understanding of a scene is important for many applications - but in what space should diverse semantic information (e.g., objects, scene categories, material types, 3D shapes, etc.) be grounded and what should be its structure? Aspiring to have one unified structure that hosts diverse types of semantics, we follow the Scene Graph paradigm in 3D, generating a 3D Scene Graph. Given a 3D mesh and registered panoramic images, we construct a graph that spans the entire building and includes semantics on objects (e.g., class, material, shape and other attributes), rooms (e.g., function, illumination type, etc.) and cameras (e.g., location, etc.), as well as the relationships among these entities. However, this process is prohibitively labor heavy if done manually. To alleviate this we devise a semi-automatic framework that employs existing detection methods and enhances them using two main constraints: I. framing of query images sampled on panoramas to maximize the performance of 2D detectors, and II. multi-view consistency enforcement across 2D detections that originate in different camera locations.

A complex visual navigation task puts an agent in different situations which call for a diverse range of visual perception abilities. For example, to "go to the nearest chair'', the agent might need to identify a chair in a living room using semantics, follow along a hallway using vanishing point cues, and avoid obstacles using depth. Therefore, utilizing the appropriate visual perception abilities based on a situational understanding of the visual environment can empower these navigation models in unseen visual environments. We propose to train an agent to fuse a large set of visual representations that correspond to diverse visual perception abilities. To fully utilize each representation, we develop an action-level representation fusion scheme, which predicts an action candidate from each representation and adaptively consolidate these action candidates into the final action. Furthermore, we employ a data-driven inter-task affinity regularization to reduce redundancies and improve generalization. Our approach leads to a significantly improved performance in novel environments over ImageNet-pretrained baseline and other fusion methods.

When operating in unstructured environments such as warehouses, homes, and retail centers, robots are frequently required to interactively search for and retrieve specific objects from cluttered bins, shelves, or tables. Mechanical Search describes the class of tasks where the goal is to locate and extract a known target object. In this paper, we formalize Mechanical Search and study a version where distractor objects are heaped over the target object in a bin. The robot uses an RGBD perception system and control policies to iteratively select, parameterize, and perform one of 3 actions - push, suction, grasp - until the target object is extracted, or either a time limit is exceeded, or no high confidence push or grasp is available. We present a study of 5 algorithmic policies for mechanical search, with 15,000 simulated trials and 300 physical trials for heaps ranging from 10 to 20 objects. Results suggest that success can be achieved in this long-horizon task with algorithmic policies in over 95% of instances and that the number of actions required scales approximately linearly with the size of the heap. Code and supplementary material can be found at http://ai.stanford.edu/mech-search.

Contact-rich manipulation tasks in unstructured environments often require both haptic and visual feedback. However, it is non-trivial to manually design a robot controller that combines modalities with very different characteristics. While deep reinforcement learning has shown success in learning control policies for high-dimensional inputs, these algorithms are generally intractable to deploy on real robots due to sample complexity. We use self-supervision to learn a compact and multimodal representation of our sensory inputs, which can then be used to improve the sample efficiency of our policy learning. We evaluate our method on a peg insertion task, generalizing over different geometry, configurations, and clearances, while being robust to external perturbations. We present results in simulation and on a real robot.

We present a navigation system that combines ideas from hierarchical planning and machine learning. The system uses a traditional global planner to compute optimal paths towards a goal, and a deep local trajectory planner and velocity controller to compute motion commands. The latter components of the system adjust the behavior of the robot through attention mechanisms such that it moves towards the goal, avoids obstacles, and respects the space of nearby pedestrians. Both the structure of the proposed deep models and the use of attention mechanisms make the system's execution interpretable. Our simulation experiments suggest that the proposed architecture outperforms baselines that try to map global plan information and sensor data directly to velocity commands. In comparison to a hand-designed traditional navigation system, the proposed approach showed more consistent performance.

We address one-shot imitation learning, where the goal is to execute a previously unseen task based on a single demonstration. While there has been exciting progress in this direction, most of the approaches still require a few hundred tasks for meta-training, which limits the scalability of the approaches. Our main contribution is to formulate one-shot imitation learning as a symbolic planning problem along with the symbol grounding problem. This formulation disentangles the policy execution from the inter-task generalization and leads to better data efficiency. The key technical challenge is that the symbol grounding is prone to error with limited training data and leads to subsequent symbolic planning failures. We address this challenge by proposing a continuous relaxation of the discrete symbolic planner that directly plans on the probabilistic outputs of the symbol grounding model. Our continuous relaxation of the planner can still leverage the information contained in the probabilistic symbol grounding and significantly improve over the baseline planner for the one-shot imitation learning tasks without using large training data.

Large, richly annotated datasets have accelerated progress in fields such as computer vision and natural language processing, but replicating these successes in robotics has been challenging. While prior data collection methodologies such as self-supervision have resulted in large datasets, the data can have poor signal-to-noise ratio. By contrast, previous efforts to collect task demonstrations with humans provide better quality data, but they cannot reach the same data magnitude. Furthermore, neither approach places guarantees on the diversity of the data collected, in terms of solution strategies. In this work, we leverage and extend the RoboTurk platform to scale up data collection for robotic manipulation using remote teleoperation. The primary motivation for our platform is two-fold: (1) to address the shortcomings of prior work and increase the total quantity of manipulation data collected through human supervision by an order of magnitude without sacrificing the quality of the data and (2) to collect data on challenging manipulation tasks across several operators and observe a diverse set of emergent behaviors and solutions. We collected over 111 hours of robot manipulation data across 54 users and 3 challenging manipulation tasks in 1 week, resulting in the largest robot dataset collected via remote teleoperation. We evaluate the quality of our platform, the diversity of demonstrations in our dataset, and the utility of our dataset via quantitative and qualitative analysis. For additional results, supplementary videos, and to download our dataset, visit http;//roboturk.stanford.edu/realrobotdataset.

Reinforcement Learning (RL) of contact-rich manipulation tasks has yielded impressive results in recent years. While many studies in RL focus on varying the observation space or reward model, few efforts focused on the choice of action space (e.g. joint or end-effector space, position, velocity, etc.). However, studies in robot motion control indicate that choosing an action space that conforms to the characteristics of the task can simplify exploration and improve robustness to disturbances. This paper studies the effect of different action spaces in deep RL and advocates for variable impedance control in end-effector space (VICES) as an advantageous action space for constrained and contact-rich tasks. We evaluate multiple action spaces on three prototypical manipulation tasks: Path Following (task with no contact), Door Opening (task with kinematic constraints), and Surface Wiping (task with continuous contact). We show that VICES improves sample efficiency, maintains low energy consumption, and ensures safety across all three experimental setups. Further, RL policies learned with VICES can transfer across different robot models in simulation, and from simulation to real for the same robot. Further information is available at https://stanfordvl.github.io/vices.

Predicting the future trajectories of multiple interacting pedestrians in a scene has become an increasingly important problem for many different applications ranging from control of autonomous vehicles and social robots to security and surveillance. This problem is compounded by the presence of social interactions between humans and their physical interactions with the scene. While the existing literature has explored some of these cues, they mainly ignored the multimodal nature of each human's future trajectory which is noticeably influenced by the intricate social interactions. In this paper, we present Social-BiGAT, a graph-based generative adversarial network that generates realistic, multimodal trajectory predictions for multiple pedestrians in a scene. Our method is based on a graph attention network (GAT) that learns feature representations that encode the social interactions between humans in the scene, and a recurrent encoder-decoder architecture that is trained adversarially to predict, based on the features, the humans' paths. We explicitly account for the multimodal nature of the prediction problem by forming a reversible transformation between each scene and its latent noise vector, as in Bicycle-GAN. We show that our framework achieves state-of-the-art performance comparing it to several baselines on existing trajectory forecasting benchmarks.

Recent learning-to-plan methods have shown promising results on planning directly from observation space. Yet, their ability to plan for long-horizon tasks is limited by the accuracy of the prediction model. On the other hand, classical symbolic planners show remarkable capabilities in solving long-horizon tasks, but they require predefined symbolic rules and symbolic states, restricting their real-world applicability. In this work, we combine the benefits of these two paradigms and propose a learning-to-plan method that can directly generate a long-term symbolic plan conditioned on high-dimensional observations. We borrow the idea of regression (backward) planning from classical planning literature and introduce Regression Planning Networks (RPN), a neural network architecture that plans backward starting at a task goal and generates a sequence of intermediate goals that reaches the current observation. We show that our model not only inherits many favorable traits from symbolic planning --including the ability to solve previously unseen tasks-- but also can learn from visual inputs in an end-to-end manner. We evaluate the capabilities of RPN in a grid world environment and a simulated 3D kitchen environment featuring complex visual scenes and long task horizon, and show that it achieves near-optimal performance in completely new task instances.

Watching expert demonstrations is an important way for humans and robots to reason about affordances of unseen objects. In this paper, we consider the problem of reasoning object affordances through the feature embedding of demonstration videos. We design the Demo2Vec model which learns to extract embedded vectors of demonstration videos and predicts the interaction region and the action label on a target image of the same object. We introduce the Online Product Review dataset for Affordance (OPRA) by collecting and labeling diverse YouTube product review videos. Our Demo2Vec model outperforms various recurrent neural network baselines on the collected dataset.

Understanding human motion behavior is critical for autonomous moving platforms (like self-driving cars and social robots) if they are to navigate human-centric environments. This is challenging because human motion is inherently multimodal: given a history of human motion paths, there are many socially plausible ways that people could move in the future. We tackle this problem by combining tools from sequence prediction and generative adversarial networks: a recurrent sequence-to-sequence model observes motion histories and predicts future behavior, using a novel pooling mechanism to aggregate information across people. We predict socially plausible futures by training adversarially against a recurrent discriminator, and encourage diverse predictions with a novel variety loss. Through experiments on several datasets we demonstrate that our approach outperforms prior work in terms of accuracy, variety, collision avoidance, and computational complexity.

Unlike machines, humans learn through rapid, Abstract model-building. The role of a teacher is not simply to hammer home right or wrong answers, but rather to provide intuitive comments, comparisons, and explanations to a pupil. This is what the Learning Under Privileged Information (LUPI) paradigm endeavors to model by utilizing extra knowledge only available during training. We propose a new LUPI algorithm specifically designed for Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). We propose to use a heteroscedastic dropout (i.e. dropout with a varying variance) and make the variance of the dropout a function of privileged information. Intuitively, this corresponds to using the privileged information to control the uncertainty of the model output. We perform experiments using CNNs and RNNs for the tasks of image classification and machine translation. Our method significantly increases the sample efficiency during learning, resulting in higher accuracy with a large margin when the number of training examples is limited. We also theoretically justify the gains in sample efficiency by providing a generalization error bound decreasing with O(1/n), where n is the number of training examples, in an oracle case.

We present Im2Pano3D, a convolutional neural network that generates a dense prediction of 3D structure and a probability distribution of semantic labels for a full 360° panoramic view of an indoor scene when given only a partial observation (= 50%) in the form of an RGB-D image. To make this possible, Im2Pano3D leverages strong contextual priors learned from large-scale synthetic and real-world indoor scenes. To ease the prediction of 3D structure, we propose to parameterize 3D surfaces with their plane equations and train the model to predict these parameters directly. To provide meaningful training supervision, we use multiple loss functions that consider both pixel level accuracy and global context consistency. Experiments demonstrate that Im2Pano3D is able to predict the semantics and 3D structure of the unobserved scene with more than 56% pixel accuracy and less than 0.52m average distance error, which is significantly better than alternative approaches.

Recent works showed that Generative Adversarial Networks (GANs) can be successfully applied in unsupervised domain adaptation, where, given a labeled source dataset and an unlabeled target dataset, the goal is to train powerful classifiers for the target samples. In particular, it was shown that a GAN objective function can be used to learn target features indistinguishable from the source ones. In this work, we extend this framework by (i) forcing the learned feature extractor to be domain-invariant, and (ii) training it through data augmentation in the feature space, namely performing feature augmentation. While data augmentation in the image space is a well established technique in deep learning, feature augmentation has not yet received the same level of attention. We accomplish it by means of a feature generator trained by playing the GAN minimax game against source features. Results show that both enforcing domain-invariance and performing feature augmentation lead to superior or comparable performance to state-of-the-art results in several unsupervised domain adaptation benchmarks.

Developing visual perception models for active agents and sensorimotor control in the physical world are cumbersome as existing algorithms are too slow to efficiently learn in real-time and robots are fragile and costly. This has given rise to learning-in-simulation which consequently casts a question on whether the results transfer to real-world. In this paper, we investigate developing real-world perception for active agents, propose Gibson Environment for this purpose, and showcase a set of perceptual tasks learned therein. Gibson is based upon virtualizing real spaces, rather than artificially designed ones, and currently includes over 1400 floor spaces from 572 full buildings. The main characteristics of Gibson are: I. being from the real-world and reflecting its semantic complexity, II. having an internal synthesis mechanism "Goggles" enabling deploying the trained models in real-world without needing domain adaptation, III. embodiment of agents and making them subject to constraints of physics and space.

Do visual tasks have relationships, or are they unrelated? For instance, could having surface normals simplify estimating the depth of an image? Intuition answers these questions positively, implying existence of a certain structure among visual tasks. Knowing this structure has notable values; it provides a principled way for identifying relationships across tasks, for instance, in order to reuse supervision among tasks with redundancies or solve many tasks in one system without piling up the complexity. We propose a fully computational approach for modeling the transfer learning structure of the space of visual tasks. This is done via finding transfer learning dependencies across tasks in a dictionary of twenty-six 2D, 2.5D, 3D, and semantic tasks. The product is a computational taxonomic map among tasks for transfer learning, and we exploit it to reduce the demand for labeled data. For example, we show that the total number of labeled datapoints needed for solving a set of 10 tasks can be reduced by roughly 2/3 (compared to training independently) while keeping the performance nearly the same. We provide a set of tools for computing and visualizing this taxonomical structure at http://taskonomy.vision.

Xiaoxue Zang, Ashwini Pokle, Marynel Vázquez, Kevin Chen, Juan Carlos Niebles, Alvaro Soto, Silvio Savarese. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018.

Learning-based approaches to robotic manipulation are limited by the scalability of data collection and accessibility of labels. In this paper, we present a multi-task domain adaptation framework for instance grasping in cluttered scenes by utilizing simulated robot experiments. Our neural network takes monocular RGB images and the instance segmentation mask of a specified target object as inputs, and predicts the probability of successfully grasping the specified object for each candidate motor command. The proposed transfer learning framework trains a model for instance grasping in simulation and uses a domain-adversarial loss to transfer the trained model to real robots using indiscriminate grasping data, which is available both in simulation and the real world. We evaluate our model in real-world robot experiments, comparing it with alternative model architectures as well as an indiscriminate grasping baseline.

In this work, we propose a novel robot learning framework called Neural Task Programming (NTP), which bridges the idea of few-shot learning from demonstration and neural program induction. NTP takes as input a task specification (e.g., video demonstration of a task) and recursively decomposes it into finer sub-task specifications. These specifications are fed to a hierarchical neural program, where bottom-level programs are callable subroutines that interact with the environment. We validate our method in three robot manipulation tasks. NTP achieves strong generalization across sequential tasks that exhibit hierarchal and compositional structures. The experimental results show that NTP learns to generalize well towards unseen tasks with increasing lengths, variable topologies, and changing objectives.stanfordvl.github.io/ntp/.

We present semi-supervised deep learning approaches for traversability estimation from fisheye images. Our method, GONet, and the proposed extensions leverage Generative Adversarial Networks (GANs) to effectively predict whether the area seen in the input image(s) is safe for a robot to traverse. These methods are trained with many positive images of traversable places, but just a small set of negative images depicting blocked and unsafe areas. This makes the proposed methods practical. Positive examples can be collected easily by simply operating a robot through traversable spaces, while obtaining negative examples is time consuming, costly, and potentially dangerous. Through extensive experiments and several demonstrations, we show that the proposed traversability estimation approaches are robust and can generalize to unseen scenarios. Further, we demonstrate that our methods are memory efficient and fast, allowing for real-time operation on a mobile robot with single or stereo fisheye cameras. As part of our contributions, we open-source two new datasets for traversability estimation. These datasets are composed of approximately 24h of videos from more than 25 indoor environments. Our methods outperform baseline approaches for traversability estimation on these new datasets.

We are concerned with learning models that generalize well to different unseen domains. We consider a worst-case formulation over data distributions that are near the source domain in the feature space. Only using training data from a single source distribution, we propose an iterative procedure that augments the dataset with examples from a fictitious target domain that is "hard" under the current model. We show that our iterative scheme is an adaptive data augmentation method where we append adversarial examples at each iteration. For softmax losses, we show that our method is a data-dependent regularization scheme that behaves differently from classical regularizers that regularize towards zero (e.g., ridge or lasso). On digit recognition and semantic segmentation tasks, our method learns models improve performance across a range of a priori unknown target domains.

The main challenge of online multi-object tracking is to reliably associate object trajectories with detections in each video frame based on their tracking history. In this work, we propose the Recurrent Autoregressive Network (RAN), a temporal generative modeling framework to characterize the appearance and motion dynamics of multiple objects over time. The RAN couples an external memory and an internal memory. The external memory explicitly stores previous inputs of each trajectory in a time window, while the internal memory learns to summarize long-term tracking history and associate detections by processing the external memory. We conduct experiments on the MOT 2015 and 2016 datasets to demonstrate the robustness of our tracking method in highly crowded and occluded scenes. Our method achieves top-ranked results on the two benchmarks.

3D reconstruction from a single image is a key problem in multiple applications ranging from robotic manipulation to augmented reality. Prior methods have tackled this problem through generative models which predict 3D reconstructions as voxels or point clouds. However, these methods can be computationally expensive and miss fine details. We introduce a new differentiable layer for 3D data deformation and use it in DEFORMNET to learn a model for 3D reconstruction-through-deformation. DEFORMNET takes an image input, finds a nearest shape template from a database, and deforms the template to match the query image. We evaluate our approach on the ShapeNet dataset and show that - (a) the Free-Form Deformation layer is a powerful new building block for Deep Learning models that manipulate 3D data (b) DEFORMNET uses this FFD layer combined with shape retrieval for smooth and detail-preserving 3D reconstruction of qualitatively plausible point clouds with respect to a single query image (c) compared to other state-of-the-art 3D reconstruction methods, DEFORMNET quantitatively matches or outperforms their benchmarks by significant margins.

There is a large variation in the activities that humans perform in their everyday lives. We consider modeling these composite human activities which comprises multiple basic level actions in a completely unsupervised setting. Our model learns high-level co-occurrence and temporal relations between the actions. We consider the video as a sequence of short-term action clips, which contains human-words and object-words. An activity is about a set of action-topics and object-topics indicating which actions are present and which objects are interacting with. We then propose a new probabilistic model relating the words and the topics. It allows us to model long-range action relations that commonly exist in the composite activities, which is challenging in previous works. We apply our model to the unsupervised action segmentation and clustering, and to a novel application that detects forgotten actions, which we call action patching. For evaluation, we contribute a new challenging RGB-D activity video dataset recorded by the new Kinect v2, which contains several human daily activities as compositions of multiple actions interacting with different objects. Moreover, we develop a robotic system that watches and reminds people using our action patching algorithm. Our robotic setup can be easily deployed on any assistive robots.

Supervised 3D reconstruction has witnessed a significant progress through the use of deep neural networks. However, this increase in performance requires large scale annotations of 2D/3D data. In this paper, we explore inexpensive 2D supervision as an alternative for expensive 3D CAD annotation. Specifically, we use foreground masks as weak supervision through a raytrace pooling layer that enables perspective projection and backpropagation. Additionally, since the 3D reconstruction from masks is an ill posed problem, we propose to constrain the 3D reconstruction to the manifold of unlabeled realistic 3D shapes that match mask observations. We demonstrate that learning a log-barrier solution to this constrained optimization problem resembles the GAN objective, enabling the use of existing tools for training GANs. We evaluate and analyze the manifold constrained reconstruction on various datasets for single and multi-view reconstruction of both synthetic and real images.

In this paper we focus on the problem of inferring geometric and semantic properties of a complex scene where humans interact with objects from egocentric views. Unlike most previous work, our goal is to leverage a multimodal sensory stream composed of RGB, depth, and thermal (RGB-D-T) signals and use this data stream as an input to a new framework for joint 6 DOF camera localization, 3D reconstruction, and semantic segmentation. As our extensive experimental evaluation shows, the combination of different sensing modalities allows us to achieve greater robustness in situations where both the observer and the objects in the scene move rapidly (a challenging situation for traditional semantic reconstruction methods). Moreover, we contribute a new dataset that includes a large number of egocentric RGB-D-T videos of humans performing daily real-world activities as well as a new demonstration hardware platform for acquiring such a dataset.

3D semantic scene labeling is fundamental to agents operating in the real world. In particular, labeling raw 3D point sets from sensors provides fine-grained semantics. Recent works leverage the capabilities of Neural Networks(NNs), but are limited to coarse voxel predictions and do not explicitly enforce global consistency. We present SEGCloud, an end-to-end framework to obtain 3D point-level segmentation that combines the advantages of NNs, trilinear interpolation(TI) and fully connected Conditional Random Fields (FC-CRF). Coarse voxel predictions from a 3D Fully Convolutional NN are transferred back to the raw 3D points via trilinear interpolation. Then the FC-CRF enforces global consistency and provides fine-grained semantics on the points. We implement the latter as a differentiable Recurrent NN to allow joint optimization. We evaluate the framework on two indoor and two outdoor 3D datasets (NYU V2, S3DIS, KITTI, Semantic3D.net), and show performance comparable or superior to the state-of-the-art on all datasets.

We present a unified framework for understanding human social behaviors in raw image sequences. Our model jointly detects multiple individuals, infers their social actions, and estimates the collective actions with a single feed-forward pass through a neural network. We propose a single architecture that does not rely on external detection algorithms but rather is trained end-to-end to generate dense proposal maps that are refined via a novel inference scheme. The temporal consistency is handled via a person-level matching Recurrent Neural Network. The complete model takes as input a sequence of frames and outputs detections along with the estimates of individual actions and collective activities. We demonstrate state-of-the-art performance of our algorithm on multiple publicly available benchmarks.

Recently, convolutional neural networks (CNN) have been successfully applied to view synthesis problems. However, such CNN-based methods can suffer from lack of texture details, shape distortions, or high computational complexity. In this paper, we propose a novel CNN architecture for view synthesis called Deep View Morphing that does not suffer from these issues. To synthesize a middle view of two input images, a rectification network first rectifies the two input images. An encoder-decoder network then generates dense correspondences between the rectified images and blending masks to predict the visibility of pixels of the rectified images in the middle view. A view morphing network finally synthesizes the middle view using the dense correspondences and blending masks. We experimentally show the proposed method significantly outperforms the state-of-the-art CNN-based view synthesis method.

Urrently, the most successful learning models in computer vision are based on learning successive representations followed by a decision layer. This is usually actualized through feedforward multilayer neural networks, e.g. ConvNets, where each layer forms one of such successive representations. However, an alternative that can achieve the same goal is a feedback based approach in which the representation is formed in an iterative manner based on a feedback received from previous iterations output. We establish that a feedback based approach has several core advantages over feedforward: it enables making early predictions at the query time, its output naturally conforms to a hierarchical structure in the label space (e.g. a taxonomy), and it provides a new basis for Curriculum Learning. We observe that feedback develops a considerably different representation compared to feedforward counterparts, in line with the aforementioned advantages. We provide a general feedback based learning architecture, instantiated using existing RNNs, with the endpoint results on par or better than existing feedforward networks and the addition of the above advantages.

The majority of existing solutions to the Multi-Target Tracking (MTT) problem do not combine cues over a long period of time in a coherent fashion. In this paper, we present an online method that encodes long-term temporal dependencies across multiple cues. One key challenge of tracking methods is to accurately track occluded targets or those which share similar appearance properties with surrounding objects. To address this challenge, we present a structure of Recurrent Neural Networks (RNN) that jointly reasons on multiple cues over a temporal window. Our method allows to correct data association errors and recover observations from occluded states. We demonstrate the robustness of our data-driven approach by tracking multiple targets using their appearance, motion, and even interactions. Our method outperforms previous works on multiple publicly available datasets including the challenging MOT benchmark.

Human actions captured in video sequences are threedimensional signals characterizing visual appearance and motion dynamics. To learn action patterns, existing methods adopt Convolutional and/or Recurrent Neural Networks (CNNs and RNNs). CNN based methods are effective in learning spatial appearances, but are limited in modeling long-term motion dynamics. RNNs, especially Long Short- Term Memory (LSTM), are able to learn temporal motion dynamics. However, naively applying RNNs to video sequences in a convolutional manner implicitly assumes that motions in videos are stationary across different spatial locations. This assumption is valid for short-term motions but invalid when the duration of the motion is long.,,In this work, we propose Lattice-LSTM (L2STM), which extends LSTM by learning independent hidden state transitions of memory cells for individual spatial locations. This method effectively enhances the ability to model dynamics across time and addresses the non-stationary issue of long-term motion dynamics without significantly increasing the model complexity. Additionally, we introduce a novel multi-modal training procedure for training our network. Unlike traditional two-stream architectures which use RGB and optical flow information as input, our two-stream model leverages both modalities to jointly train both input gates and both forget gates in the network rather than treating the two streams as separate entities with no information about the other. We apply this end-to-end system to benchmark datasets (UCF-101 and HMDB-51) of human action recognition. Experiments show that on both datasets, our proposed method outperforms all existing ones that are based on LSTM and/or CNNs of similar model complexities.

Existing camera networks in public spaces such as train terminals or malls can help social robots to navigate crowded scenes. However, the localization of the cameras is required, i.e., the positions and poses of all cameras in a unique reference. In this work, we estimate the relative location of any pair of cameras by solely using noisy trajectories observed from each camera. We propose a fully unsupervised learning technique using unlabelled pedestrians motion patterns captured in crowded scenes. We first estimate the pairwise camera parameters by optimally matching single-view pedestrian tracks using social awareness. Then, we show the impact of jointly estimating the network parameters. This is done by formulating a nonlinear least square optimization problem, leveraging a continuous approximation of the matching function. We evaluate our approach in real-world environments such as train terminals, where several hundreds of individuals need to be tracked across dozens of cameras every second.

Policy search methods in reinforcement learning have demonstrated success in scaling up to larger problems beyond toy examples. However, deploying these methods on real robots remains challenging due to the large sample complexity required during learning and their vulnerability to malicious intervention. We introduce Adversarially Robust Policy Learning (ARPL), an algorithm that leverages active computation of physically-plausible adversarial examples during training to enable robust policy learning in the source domain and robust performance under both random and adversarial input perturbations. We evaluate ARPL on four continuous control tasks and show superior resilience to changes in physical environment dynamics parameters and environment state as compared to state-of-the-art robust policy learning methods. Code, data, and additional experimental results are available at: stanfordvl.github.io/ARPL.

In Convolutional Neural Network (CNN)-based object detection methods, region proposal becomes a bottleneck when objects exhibit significant scale variation, occlusion or truncation. In addition, these methods mainly focus on 2D object detection and cannot estimate detailed properties of objects. In this paper, we propose subcategory-aware CNNs for object detection. We introduce a novel region proposal network that uses subcategory information to guide the proposal generating process, and a new detection network for joint detection and subcategory classification. By using subcategories related to object pose, we achieve state of-the-art performance on both detection and pose estimation on commonly used benchmarks.

Pedestrians follow different trajectories to avoid obstacles and accommodate fellow pedestrians. Any autonomous vehicle navigating such a scene should be able to foresee the future positions of pedestrians and accordingly adjust its path to avoid collisions. This problem of trajectory prediction can be viewed as a sequence generation task, where we are interested in predicting the future trajectory of people based on their past positions. Following the recent success of Recurrent Neural Network (RNN) models for sequence prediction tasks, we propose an LSTM model which can learn general human movement and predict their future trajectories. This is in contrast to traditional approaches which use hand-crafted functions such as Social forces. We demonstrate the performance of our method on several public datasets. Our model outperforms state-of-the-art methods on some of these datasets. We also analyze the trajectories predicted by our model to demonstrate the motion behaviour learned by our model.

In this paper, we propose a method for semantic parsing the 3D point cloud of an entire building using a hierarchical approach: first, the raw data is parsed into semantically meaningful spaces (e.g. rooms, etc) that are aligned into a canonical reference coordinate system. Second, the spaces are parsed into their structural and building elements (e.g. walls, columns, etc). Performing these with a strong notation of global 3D space is the backbone of our method. The alignment in the first step injects strong 3D priors from the canonical coordinate system into the second step for discovering elements. This allows diverse challenging scenarios as man-made indoor spaces often show recurrent geometric patterns while the appearance features can change drastically. We also argue that identification of structural elements in indoor spaces is essentially a detection problem, rather than segmentation which is commonly used. We evaluated our method on a new dataset of several buildings with a covered area of over 6, 000m2and over 215 million points, demonstrating robust results readily useful for practical applications.

We consider the problem of estimating the spatial layout of an indoor scene from a monocular RGB image, modeled as the projection of a 3D cuboid. Existing solutions to this problem often rely strongly on hand-engineered features and vanishing point detection, which are prone to failure in the presence of clutter. In this paper, we present a method that uses a fully convolutional neural network (FCNN) in conjunction with a novel optimization framework for generating layout estimates. We demonstrate that our method is robust in the presence of clutter and handles a wide range of highly challenging scenes. We evaluate our method on two standard benchmarks and show that it achieves state of the art results, outperforming previous methods by a wide margin.

Deep Recurrent Neural Network architectures, though remarkably capable at modeling sequences, lack an intuitive high-level spatio-temporal structure. That is while many problems in computer vision inherently have an underlying high-level structure and can benefit from it. Spatiotemporal graphs are a popular tool for imposing such high-level intuitions in the formulation of real world problems. In this paper, we propose an approach for combining the power of high-level spatio-temporal graphs and sequence learning success of Recurrent Neural Networks (RNNs). We develop a scalable method for casting an arbitrary spatio-temporal graph as a rich RNN mixture that is feedforward, fully differentiable, and jointly trainable. The proposed method is generic and principled as it can be used for transforming any spatio-temporal graph through employing a certain set of well defined steps. The evaluations of the proposed approach on a diverse set of problems, ranging from modeling human motion to object interactions, shows improvement over the state-of-the-art with a large margin. We expect this method to empower new approaches to problem formulation through high-level spatio-temporal graphs and Recurrent Neural Networks.

Learning the distance metric between pairs of examples is of great importance for learning and visual recognition. With the remarkable success from the state of the art convolutional neural networks, recent works [1, 31] have shown promising results on discriminatively training the networks to learn semantic feature embeddings where similar examples are mapped close to each other and dissimilar examples are mapped farther apart. In this paper, we describe an algorithm for taking full advantage of the training batches in the neural network training by lifting the vector of pairwise distances within the batch to the matrix of pairwise distances. This step enables the algorithm to learn the state of the art feature embedding by optimizing a novel structured prediction objective on the lifted problem. Additionally, we collected Stanford Online Products dataset: 120k images of 23k classes of online products for metric learning. Our experiments on the CUB-200-2011 [37], CARS196 [19], and Stanford Online Products datasets demonstrate significant improvement over existing deep feature embedding methods on all experimented embedding sizes with the GoogLeNet [33] network. The source code and the dataset are available at: https://github.com/rksltnl/ Deep-Metric-Learning-CVPR16.

Some robots must repeatedly interact with a fixed set of objects in their environment. To operate correctly, it is helpful for the robot to be able to recognize the object instances that it repeatedly encounters. However, current methods for recognizing object instances require that, during training, many pictures are taken of each object from a large number of viewing angles. This procedure is slow and requires much manual effort before the robot can begin to operate in a new environment. We have developed a novel procedure for training a neural network to recognize a set of objects from just a single training image per object. To obtain robustness to changes in viewpoint, we take advantage of a supplementary dataset in which we observe a separate (non-overlapping) set of objects from multiple viewpoints. After pre-training the network in a novel multi-stage fashion, the network can robustly recognize new object instances given just a single training image of each object. If more images of each object are available, the performance improves. We perform a thorough analysis comparing our novel training procedure to traditional neural network pre-training techniques as well as previous state-of-the-art approaches including keypoint-matching, template-matching, and sparse coding, and we demonstrate that our method significantly outperforms these previous approaches. Our method can thus be used to easily teach a robot to recognize a novel set of object instances from unknown viewpoints.

We present a robotic system that watches a human using a Kinect v2 RGB-D sensor, detects what he forgot to do while performing an activity, and if necessary reminds the person using a laser pointer to point out the related object. Our simple setup can be easily deployed on any assistive robot. Our approach is based on a learning algorithm trained in a purely unsupervised setting, which does not require any human annotations. This makes our approach scalable and applicable to variant scenarios. Our model learns the action/object co-occurrence and action temporal relations in the activity, and uses the learned rich relationships to infer the forgotten action and the related object. We show that our approach not only improves the unsupervised action segmentation and action cluster assignment performance, but also effectively detects the forgotten actions on a challenging human activity RGB-D video dataset. In robotic experiments, we show that our robot is able to remind people of forgotten actions successfully.

We present a deep learning framework for accurate visual correspondences and demonstrate its effectiveness for both geometric and semantic matching, spanning across rigid motions to intra-class shape or appearance variations. In contrast to previous CNN-based approaches that optimize a surrogate patch similarity objective, we use deep metric learning to directly learn a feature space that preserves either geometric or semantic similarity. Our fully convolutional architecture, along with a novel correspondence contrastive loss allows faster training by effective reuse of computations, accurate gradient computation through the use of thousands of examples per image pair and faster testing with $O(n)$ feedforward passes for n keypoints, instead of $O(n^2)$ for typical patch similarity methods. We propose a convolutional spatial transformer to mimic patch normalization in traditional features like SIFT, which is shown to dramatically boost accuracy for semantic correspondences across intra-class shape variations. Extensive experiments on KITTI, PASCAL and CUB-2011 datasets demonstrate the significant advantages of our features over prior works that use either hand-constructed or learned features.

Supervised learning with large scale labelled datasets and deep layered models has caused a paradigm shift in diverse areas in learning and recognition. However, this approach still suffers from generalization issues under the presence of a domain shift between the training and the test data distribution. Since unsupervised domain adaptation algorithms directly address this domain shift problem between a labelled source dataset and an unlabelled target dataset, recent papers have shown promising results by fine-tuning the networks with domain adaptation loss functions which try to align the mismatch between the training and testing data distributions. Nevertheless, these recent deep learning based domain adaptation approaches still suffer from issues such as high sensitivity to the gradient reversal hyperparameters and overfitting during the fine-tuning stage. In this paper, we propose a unified deep learning framework where the representation, cross domain transformation, and target label inference are all jointly optimized in an end-to-end fashion for unsupervised domain adaptation. Our experiments show that the proposed method significantly outperforms state-of-the-art algorithms in both object recognition and digit classification experiments by a large margin. We will make our learned models as well as the source code available immediately upon acceptance.

A large body of recent work on object detection has focused on exploiting 3D CAD model databases to improve detection performance. Many of these approaches work by aligning exact 3D models to images using templates generated from renderings of the 3D models at a set of discrete viewpoints. However, the training procedures for these approaches are computationally expensive and require gigabytes of memory and storage, while the viewpoint discretization hampers pose estimation performance. We propose an efficient method for synthesizing templates from 3D models that runs on the fly - that is, it quickly produces detectors for an arbitrary viewpoint of a 3D model without expensive dataset-dependent training or template storage. Given a 3D model and an arbitrary continuous detection viewpoint, our method synthesizes a discriminative template by extracting features from a rendered view of the object and decorrelating spatial dependences among the features. Our decorrelation procedure relies on a gradient-based algorithm that is more numerically stable than standard decomposition-based procedures, and we efficiently search for candidate detections by computing FFT-based template convolutions. Due to the speed of our template synthesis procedure, we are able to perform joint optimization of scale, translation, continuous rotation, and focal length using Metropolis-Hastings algorithm. We provide an efficient GPU implementation of our algorithm, and we validate its performance on 3D Object Classes and PASCAL3D+ datasets.

Despite the fact that object detection, 3D pose estimation, and sub-category recognition are highly correlated tasks, they are usually addressed independently from each other because of the huge space of parameters. To jointly model all of these tasks, we propose a coarse-to-fine hierarchical representation, where each level of the hierarchy represents objects at a different level of granularity. The hierarchical representation prevents performance loss, which is often caused by the increase in the number of parameters (as we consider more tasks to model), and the joint modeling enables resolving ambiguities that exist in independent modeling of these tasks. We augment PASCAL3D+ [34] dataset with annotations for these tasks and show that our hierarchical model is effective in joint modeling of object detection, 3D pose estimation, and sub-category recognition.

We focus on modeling human activities comprising multiple actions in a completely unsupervised setting. Our model learns the high-level action co-occurrence and temporal relations between the actions in the activity video. We consider the video as a sequence of short-term action clips, called action-words, and an activity is about a set of action-topics indicating which actions are present in the video. Then we propose a new probabilistic model relating the action-words and the action-topics. It allows us to model long-range action relations that commonly exist in the complex activity, which is challenging to capture in the previous works. We apply our model to unsupervised action segmentation and recognition, and also to a novel application that detects forgotten actions, which we call action patching. For evaluation, we also contribute a new challenging RGB-D activity video dataset recorded by the new Kinect v2, which contains several human daily activities as compositions of multiple actions interacted with different objects. The extensive experiments show the effectiveness of our model.

Despite the great progress achieved in recognizing objects as 2D bounding boxes in images, it is still very challenging to detect occluded objects and estimate the 3D properties of multiple objects from a single image. In this paper, we propose a novel object representation, 3D Voxel Pattern (3DVP), that jointly encodes the key properties of objects including appearance, 3D shape, viewpoint, occlusion and truncation. We discover 3DVPs in a data-driven way, and train a bank of specialized detectors for a dictionary of 3DVPs. The 3DVP detectors are capable of detecting objects with specific visibility patterns and transferring the meta-data from the 3DVPs to the detected objects, such as 2D segmentation mask, 3D pose as well as occlusion or truncation boundaries. The transferred meta-data allows us to infer the occlusion relationship among objects, which in turn provides improved object recognition results. Experiments are conducted on the KITTI detection benchmark [17] and the outdoor-scene dataset [41]. We improve state-of-the-art results on car detection and pose estimation with notable margins (6% in difficult data of KITTI). We also verify the ability of our method in accurately segmenting objects from the background and localizing them in 3D.

Realistic videos of human actions exhibit rich spatiotemporal structures at multiple levels of granularity: an action can always be decomposed into multiple finer-grained elements in both space and time. To capture this intuition, we propose to represent videos by a hierarchy of mid-level action elements (MAEs), where each MAE corresponds to an action-related spatiotemporal segment in the video. We introduce an unsupervised method to generate this representation from videos. Our method is capable of distinguishing action-related segments from background segments and representing actions at multiple spatiotemporal resolutions. Given a set of spatiotemporal segments generated from the training data, we introduce a discriminative clustering algorithm that automatically discovers MAEs at multiple levels of granularity. We develop structured models that capture a rich set of spatial, temporal and hierarchical relations among the segments, where the action label and multiple levels of MAE labels are jointly inferred. The proposed model achieves state-of-the-art performance in multiple action recognition benchmarks. Moreover, we demonstrate the effectiveness of our model in real-world applications such as action recognition in large-scale untrimmed videos and action parsing.

Human communication typically has an underlying structure. This is reflected in the fact that in many user generated videos, a starting point, ending, and certain objective steps between these two can be identified. In this paper, we propose a method for parsing a video into such semantic steps in an unsupervised way. The proposed method is capable of providing a semantic "storyline" of the video composed of its objective steps. We accomplish this utilizing both visual and language cues in a joint generative model. The proposed method can also provide a textual description for each of identified semantic steps and video segments. We evaluate this method on a large number of complex YouTube videos and show results of unprecedented quality for this new and impactful problem.

Online Multi-Object Tracking (MOT) has wide applications in time-critical video analysis scenarios, such as robot navigation and autonomous driving. In tracking-by-detection, a major challenge of online MOT is how to robustly associate noisy object detections on a new video frame with previously tracked objects. In this work, we formulate the online MOT problem as decision making in Markov Decision Processes (MDPs), where the lifetime of an object is modeled with a MDP. Learning a similarity function for data association is equivalent to learning a policy for the MDP, and the policy learning is approached in a reinforcement learning fashion which benefits from both advantages of offline-learning and online-learning for data association. Moreover, our framework can naturally handle the birth/death and appearance/disappearance of targets by treating them as state transitions in the MDP while leveraging existing online single object tracking methods. We conduct experiments on the MOT Benchmark [24] to verify the effectiveness of our method.

We present a novel method for multiple people tracking that leverages a generalized model for capturing interactions among individuals. At the core of our model lies a learned dictionary of interaction feature strings which capture relationships between the motions of targets. These feature strings, created from low-level image features, lead to a much richer representation of the physical interactions between targets compared to hand-specified social force models that previous works have introduced for tracking. One disadvantage of using social forces is that all pedestrians must be detected in order for the forces to be applied, while our method is able to encode the effect of undetected targets, making the tracker more robust to partial occlusions. The interaction feature strings are used in a Random Forest framework to track targets according to the features surrounding them. Results on six publicly available sequences show that our method outperforms state-of-the-art approaches in multiple people tracking.

The Recurrent temporal restricted Boltzmann machine (RTRBM) is a probabilistic model for temporal data, that has been shown to effectively capture both short and long-term dependencies in time-series. The topology of the RTRBM graphical model, however, assumes full connectivity between all the pairs of visible and hidden units, therefore ignoring the dependency structure between the different observations. Learning this structure has the potential to not only improve the prediction performance, but it can also reveal important patterns in the data. For example, given an econometric dataset, we could identify interesting dependencies between different market sectors; given a meteorological dataset, we could identify regional weather patterns. In this work we propose a new class of RTRBM, which explicitly uses a dependency graph to model the structure in the problem and to define the energy function. We refer to the new model as the structured RTRBM (SRTRBM). Our technique is related to methods such as graphical lasso, which are used to learn the topology of Gaussian graphical models. We also develop a spike-and-slab version of the RTRBM, and combine it with our method to learn structure in datasets with real valued observations. Our experimental results using synthetic and real datasets, demonstrate that the SRTRBM can improve the prediction performance of the RTRBM, particularly when the number of visible units is large and the size of the training set is small. It also reveals the structure underlying our benchmark datasets.

This paper reports on a novel mutual information (MI) based algorithm for robust place recognition. The proposed method provides a principled framework for fusing the complementary information obtained from 3D lidar and camera imagery for recognizing places within an a priori map of a dynamic environment. The visual appearance of the locations in the map can be significantly different due to changing weather, lighting conditions and dynamical objects present in the environment. Various 3D/2D features are extracted from the textured point clouds (scans) and each scan is represented as a collection of these features. For two scans acquired from the same location, the high value of MI between the features present in the scans indicates that the scans are captured from the same location. We use a non-parametric entropy estimator to estimate the true MI from the sparse marginal and joint histograms of the features extracted from the scans. Experimental results using seasonal datasets collected over several years are used to validate the robustness of the proposed algorithm.

We present a novel framework for robustly understanding the geometrical and semantic structure of a cluttered room from a small number of images captured from different viewpoints. The tasks we seek to address include: i) estimating the 3D layout of the room - that is, the 3D configuration of floor, walls and ceiling; ii) identifying and localizing all the foreground objects in the room. We jointly use multiview geometry constraints and image appearance to identify the best room layout configuration. Extensive experimental evaluation demonstrates that our estimation results are more complete and accurate in estimating 3D room structure and recognizing objects than alternative state-of-the-art algorithms. In addition, we show an augmented reality mobile application to highlight the high accuracy of our method, which may be beneficial to many computer vision applications.

3D object detection and pose estimation methods have become popular in recent years since they can handle ambiguities in 2D images and also provide a richer description for objects compared to 2D object detectors. However, most of the datasets for 3D recognition are limited to a small amount of images per category or are captured in controlled environments. In this paper, we contribute PASCAL3D+ dataset, which is a novel and challenging dataset for 3D object detection and pose estimation. PASCAL3D+ augments 12 rigid categories of the PASCAL VOC 2012 [4] with 3D annotations. Furthermore, more images are added for each category from ImageNet [3]. PASCAL3D+ images exhibit much more variability compared to the existing 3D datasets, and on average there are more than 3,000 object instances per category. We believe this dataset will provide a rich testbed to study 3D detection and pose estimation and will help to significantly push forward research in this area. We provide the results of variations of DPM [6] on our new dataset for object detection and viewpoint estimation in different scenarios, which can be used as baselines for the community. Our benchmark is available online at http://cvgl.stanford.edu/projects/pascal3d.

This paper presents a principled framework for analyzing collective activities at different levels of semantic granularity from videos. Our framework is capable of jointly tracking multiple individuals, recognizing activities performed by individuals in isolation (i.e., atomic activities such as walking or standing), recognizing the interactions between pairs of individuals (i.e., interaction activities) as well as understanding the activities of group of individuals (i.e., collective activities). A key property of our work is that it can coherently combine bottom-up information stemming from detections or fragments of tracks (or tracklets) with top-down evidence. Top-down evidence is provided by a newly proposed descriptor that captures the coherent behavior of groups of individuals in a spatial-temporal neighborhood of the sequence. Top-down evidence provides contextual information for establishing accurate associations between detections or tracklets across frames and, thus, for obtaining more robust tracking results. Bottom-up evidence percolates upwards so as to automatically infer collective activity labels. Experimental results on two challenging data sets demonstrate our theoretical claims and indicate that our model achieves enhances tracking results and the best collective classification results to date.

In the last few years, substantially different approaches have been adopted for segmenting and detecting "things" (object categories that have a well defined shape such as people and cars) and "stuff" (object categories which have an amorphous spatial extent such as grass and sky). While things have been typically detected by sliding window or Hough transform based methods, detection of stuff is generally formulated as a pixel or segment-wise classification problem. This paper proposes a framework for scene understanding that models both things and stuff using a common representation while preserving their distinct nature by using a property list. This representation allows us to enforce sophisticated geometric and semantic relationships between thing and stuff categories via property interactions in a single graphical model. We use the latest advances made in the field of discrete optimization to efficiently perform maximum a posteriori (MAP) inference in this model. We evaluate our method on the Stanford dataset by comparing it against state-of-the-art methods for object segmentation and detection. We also show that our method achieves competitive performances on the challenging PASCAL '09 segmentation dataset.

Many works have been presented for indoor scene understanding, yet few of them combine structural reasoning with full motion estimation in a real-time oriented approach. In this work we address the problem of estimating the 3D structural layout of complex and cluttered indoor scenes from monocular video sequences, where the observer can freely move in the surrounding space. We propose an effective probabilistic formulation that allows us to generate, evaluate and optimize layout hypotheses by integrating new image evidence as the observer moves. Compared to state-of-the-art work, our approach makes significantly less limiting hypotheses about the scene and the observer (e.g., Manhattan world assumption, known camera motion). We introduce a new challenging dataset and present an extensive experimental evaluation, which demonstrates that our formulation reaches near-real-time computation time and outperforms state-of-the-art methods while operating in significantly less constrained conditions. 1

The capabilities of mobile devices have been increasing at a momentous rate. As better processors have merged with capable cameras in mobile systems, the number of computer vision applications has grown rapidly. However, the computational and energy constraints of mobile devices have forced computer vision application developers to sacrifice accuracy for the sake of meeting timing demands. To increase the computational performance of mobile systems we present EVA. EVA is an application-specific heterogeneous multicore having a mix of computationally powerful cores with energy efficient cores. Each core of EVA has computation and memory architectural enhancements tailored to the application traits of vision codes. Using a computer vision benchmarking suite, we evaluate the efficiency and performance of a wide range of EVA designs. We show that EVA can provide speedups of over 9× that of an embedded processor while reducing energy demands by as much as 3×.

We present a dense reconstruction approach that overcomes the drawbacks of traditional multiview stereo by incorporating semantic information in the form of learned category-level shape priors and object detection. Given training data comprised of 3D scans and images of objects from various viewpoints, we learn a prior comprised of a mean shape and a set of weighted anchor points. The former captures the commonality of shapes across the category, while the latter encodes similarities between instances in the form of appearance and spatial consistency. We propose robust algorithms to match anchor points across instances that enable learning a mean shape for the category, even with large shape variations across instances. We model the shape of an object instance as a warped version of the category mean, along with instance-specific details. Given multiple images of an unseen instance, we collate information from 2D object detectors to align the structure from motion point cloud with the mean shape, which is subsequently warped and refined to approach the actual shape. Extensive experiments demonstrate that our model is general enough to learn semantic priors for different object categories, yet powerful enough to reconstruct individual shapes with large variations. Qualitative and quantitative evaluations show that our framework can produce more accurate reconstructions than alternative state-of-the-art multiview stereo systems.

Visual scene understanding is a difficult problem interleaving object detection, geometric reasoning and scene classification. We present a hierarchical scene model for learning and reasoning about complex indoor scenes which is computationally tractable, can be learned from a reasonable amount of training data, and avoids oversimplification. At the core of this approach is the 3D Geometric Phrase Model which captures the semantic and geometric relationships between objects which frequently co-occur in the same 3D spatial configuration. Experiments show that this model effectively explains scene semantics, geometry and object groupings from a single image, while also improving individual object detections.

In this paper we focus on the problem of detecting objects in 3D from RGB-D images. We propose a novel framework that explores the compatibility between segmentation hypotheses of the object in the image and the corresponding 3D map. Our framework allows to discover the optimal location of the object using a generalization of the structural latent SVM formulation in 3D as well as the definition of a new loss function defined over the 3D space in training. We evaluate our method using two existing RGB-D datasets. Extensive quantitative and qualitative experimental results show that our proposed approach outperforms state-of-the-art as methods well as a number of baseline approaches for both 3D and 2D object recognition tasks.

The use of semantic attributes in computer vision problems has been gaining increased popularity in recent years. Attributes provide an intermediate feature representation in between low-level features and the class categories, and offer several attractive properties, among which are improved learning of novel categories based on few examples, as well as allowing for zero-shot learning. However, the major caveat is that learning semantic attributes is a laborious task, requiring a significant amount of time and human intervention to provide labels. In order to address this issue, we propose a weakly supervised approach to learn mid-level features, where the only supervision is provided by the category classes of the training examples. We develop a novel extension of the restricted Boltzmann machine (RBM) with Beta-Bernoulli process priors. Unlike the standard RBM, our model uses the class labels to promote more efficient sharing of information by different categories. This tends to improve the generalization performance. By using semantic attributes for which annotations are available, we show that we can find correspondences between the mid-level features that we learn and the labeled attributes. Therefore, the mid-level features have distinct semantic characterization which is very similar to that given by the semantic attributes, even though their labeling was not used during the training process. Our experimental results in object recognition tasks show significant performance gains, outperforming methods which rely on manually labeled semantic attributes.

Scene understanding is an important yet very challenging problem in computer vision. In the past few years, researchers have taken advantage of the recent diffusion of depth-RGB (RGB-D) cameras to help simplify the problem of inferring scene semantics. However, while the added 3D geometry is certainly useful to segment out objects with different depth values, it also adds complications in that the 3D geometry is often incorrect because of noisy depth measurements and the actual 3D extent of the objects is usually unknown because of occlusions. In this paper we propose a new method that allows us to jointly refine the 3D reconstruction of the scene (raw depth values) while accurately segmenting out the objects or scene elements from the 3D reconstruction. This is achieved by introducing a new model which we called Voxel-CRF. The Voxel-CRF model is based on the idea of constructing a conditional random field over a 3D volume of interest which captures the semantic and 3D geometric relationships among different elements (voxels) of the scene. Such model allows to jointly estimate (1) a dense voxel-based 3D reconstruction and (2) the semantic labels associated with each voxel even in presence of partial occlusions using an approximate yet efficient inference strategy. We evaluated our method on the challenging NYU Depth dataset (Version 1 and 2). Experimental results show that our method achieves competitive accuracy in inferring scene semantics and visually appealing results in improving the quality of the 3D reconstruction. We also demonstrate an interesting application of object removal and scene completion from RGB-D images.

Many methods have been proposed to solve the image classification problem for a large number of categories. Among them, methods based on tree-based representations achieve good trade-off between accuracy and test time efficiency. While focusing on learning a tree-shaped hierarchy and the corresponding set of classifiers, most of them [11, 2, 14] use a greedy prediction algorithm for test time efficiency. We argue that the dramatic decrease in accuracy at high efficiency is caused by the specific design choice of the learning and greedy prediction algorithms. In this work, we propose a classifier which achieves a better trade-off between efficiency and accuracy with a given tree-shaped hierarchy. First, we convert the classification problem as finding the best path in the hierarchy, and a novel branch-and-bound-like algorithm is introduced to efficiently search for the best path. Second, we jointly train the classifiers using a novel Structured SVM (SSVM) formulation with additional bound constraints. As a result, our method achieves a significant 4.65%, 5.43%, and 4.07% (relative 24.82%, 41.64%, and 109.79%) improvement in accuracy at high efficiency compared to state-of-the-art greedy "tree-based" methods [14] on Caltech-256 [15], SUN [32] and Image Net 1K [9] dataset, respectively. Finally, we show that our branch-and-bound-like algorithm naturally ranks the paths in the hierarchy (Fig. 8) so that users can further process them.

We present an approach to multi-target tracking that has expressive potential beyond the capabilities of chain-shaped hidden Markov models, yet has significantly reduced complexity. Our framework, which we call tracking-by-selection, is similar to tracking-by-detection in that it separates the tasks of detection and tracking, but it shifts temporal reasoning from the tracking stage to the detection stage. The core feature of tracking-by-selection is that it reasons about path hypotheses that traverse the entire video instead of a chain of single-frame object hypotheses. A traditional chain-shaped tracking-by-detection model is only able to promote consistency between one frame and the next. In tracking-by-selection, path hypotheses exist across time, and encouraging long-term temporal consistency is as simple as rewarding path hypotheses with consistent image features. One additional advantage of tracking-by-selection is that it results in a dramatically simplified model that can be solved exactly. We adapt an existing tracking-by-detection model to the tracking-by-selection framework, and show improved performance on a challenging dataset.

In this paper, we present a general framework for tracking multiple, possibly interacting, people from a mobile vision platform. To determine all of the trajectories robustly and in a 3D coordinate system, we estimate both the camera's ego-motion and the people's paths within a single coherent framework. The tracking problem is framed as finding the MAP solution of a posterior probability, and is solved using the reversible jump Markov chain Monte Carlo (RJ-MCMC) particle filtering method. We evaluate our system on challenging datasets taken from moving cameras, including an outdoor street scene video dataset, as well as an indoor RGB-D dataset collected in an office. Experimental evidence shows that the proposed method can robustly estimate a camera's motion from dynamic scenes and stably track people who are moving independently or interacting.

This paper reports on a mutual information (MI) based algorithm for automatic extrinsic calibration of a 3D laser scanner and optical camera system. By using MI as the registration criterion, our method is able to work in situ without the need for any specific calibration targets, which makes it practical for in-field calibration. The calibration parameters are estimated by maximizing the mutual information obtained between the sensor-measured surface intensities. We calculate the Cramer-Rao-Lower-Bound (CRLB) and show that the sample variance of the estimated parameters empirically approaches the CRLB for a sufficient number of views. Furthermore, we compare the calibration results to independent ground-truth and observe that the mean error also empirically approaches to zero as the number of views are increased. This indicates that the proposed algorithm, in the limiting case, calculates a minimum variance unbiased (MVUB) estimate of the calibration parameters. Experimental results are presented for data collected by a vehicle mounted with a 3D laser scanner and an omnidirectional camera system.

Structure from motion (SFM) aims at jointly recovering the structure of a scene as a collection of 3D points and estimating the camera poses from a number of input images. In this paper we generalize this concept: not only do we want to recover 3D points, but also recognize and estimate the location of high level semantic scene components such as regions and objects in 3D. As a key ingredient for this joint inference problem, we seek to model various types of interactions between scene components. Such interactions help regularize our solution and obtain more accurate results than solving these problems in isolation. Experiments on public datasets demonstrate that: 1) our framework estimates camera poses more robustly than SFM algorithms that use points only; 2) our framework is capable of accurately estimating pose and location of objects, regions, and points in the 3D scene; 3) our framework recognizes objects and regions more accurately than state-of-the-art single image recognition methods.

Mobile platforms such as smart-phones and tablet computers have attained the technological capacity to perform tasks beyond their intended purposes. The steady increase of processing power has enticed researchers to attempt increasingly challenging tasks on mobile devices with appropriate modifications over their stationary counterparts. In this work we present a novel multi-frame object detection application for the mobile platform that is capable of object localization. Our work leverages the hough forest based object detector introduced by Gall et al. in [10]. In our experiments, we demonstrate that our novel, multi-frame generalization of [10] notably improves the detection performance. We test the performance of the technique in variable resolutions, the applicability to several object categories and different datasets. We implement the multi-frame detector on a mobile platform through a novel client-server framework that presents a sound and viable environment for the multi-frame detector. Finally, we study implementations of both single and multi-frame object detectors based on this client-server framework on a mobile device running the android OS.

Human pose estimation in a static image is a challenging problem in computer vision in that body part configurations are often subject to severe deformations and occlusions. Moreover, efficient pose estimation is often a desirable requirement in many applications. The trade-off between accuracy and efficiency has been explored in a large number of approaches. On the one hand, models with simple representations (like tree or star models) can be efficiently applied in pose estimation problems. However, these models are often prone to body part misclassification errors. On the other hand, models with rich representations (i.e., loopy graphical models) are theoretically more robust, but their inference complexity may increase dramatically. In this work, we propose an efficient and exact inference algorithm based on branch-and-bound to solve the human pose estimation problem on loopy graphical models. We show that our method is empirically much faster (about 74 times) than the state-of-the-art exact inference algorithm [21]. By extending a state-of-the-art tree model [16] to a loopy graphical model, we show that the estimation accuracy improves for most of the body parts (especially lower arms) on popular datasets such as Buffy [7] and Stickmen [5] datasets. Finally, our method can be used to exactly solve most of the inference problems on Stretchable Models [18] (which contains a few hundreds of variables) in just a few minutes.

In this work we seek to move away from the traditional paradigm for 2D object recognition whereby objects are identified in the image as 2D bounding boxes. We focus instead on: i) detecting objects; ii) identifying their 3D poses; iii) characterizing the geometrical and topological properties of the objects in terms of their aspect configurations in 3D. We call such characterization an object's aspect layout (see Fig. 1). We propose a new model for solving these problems in a joint fashion from a single image for object categories. Our model is constructed upon a novel framework based on conditional random fields with maximal margin parameter estimation. Extensive experiments are conducted to evaluate our model's performance in determining object pose and layout from images. We achieve superior viewpoint accuracy results on three public datasets and show extensive quantitative analysis to demonstrate the ability of accurately recovering the aspect layout of objects.

This paper reports a novel mutual information (MI) based algorithm for automatic registration of unstructured 3D point clouds comprised of co-registered 3D lidar and camera imagery. The proposed method provides a robust and principled framework for fusing the complementary information obtained from these two different sensing modalities. High-dimensional features are extracted from a training set of textured point clouds (scans) and hierarchical k-means clustering is used to quantize these features into a set of codewords. Using this codebook, any new scan can be represented as a collection of codewords. Under the correct rigid-body transformation aligning two overlapping scans, the MI between the codewords present in the scans is maximized. We apply a James-Stein-type shrinkage estimator to estimate the true MI from the marginal and joint histograms of the codewords extracted from the scans. Experimental results using scans obtained by a vehicle equipped with a 3D laser scanner and an omnidirectional camera are used to validate the robustness of the proposed algorithm over a wide range of initial conditions. We also show that the proposed method works well with 3D data alone.

We propose a novel framework for multimodal video indexing and retrieval using shrinkage optimized directed information assessment (SODA) as similarity measure. The directed information (DI) is a variant of the classical mutual information which attempts to capture the direction of information flow that videos naturally possess. It is applied directly to the empirical probability distributions of both audio-visual features over successive frames. We utilize RASTA-PLP features for audio feature representation and SIFT features for visual feature representation. We compute the joint probability density functions of audio and visual features in order to fuse features from different modalities. With SODA, we further estimate the DI in a manner that is suitable for high dimensional featurespand small sample sizen(largepsmalln) between pairs of video-audio modalities. We demonstrate the superiority of the SODA approach in video indexing, retrieval, and activity recognition as compared to the state-of-the-art methods such as hidden Markov models (HMM), support vector machine (SVM), cross-media indexing space (CMIS), and other noncausal divergence measures such as mutual information (MI). We also demonstrate the success of SODA in audio and video localization and indexing/retrieval of data with missaligned modalities.

Conventional rigid structure from motion (SFM) addresses the problem of recovering the camera parameters (motion) and the 3D locations (structure) of scene points, given observed 2D image feature points. In this paper, we propose a new formulation called Semantic Structure From Motion (SSFM). In addition to the geometrical constraints provided by SFM, SSFM takes advantage of both semantic and geometrical properties associated with objects in the scene (Fig. 1). These properties allow us to recover not only the structure and motion but also the 3D locations, poses, and categories of objects in the scene. We cast this problem as a max-likelihood problem where geometry (cameras, points, objects) and semantic information (object classes) are simultaneously estimated. The key intuition is that, in addition to image features, the measurements of objects across views provide additional geometrical constraints that relate cameras and scene parameters. These constraints make the geometry estimation process more robust and, in turn, make object detection more accurate. Our framework has the unique ability to: i) estimate camera poses only from object detections, ii) enhance camera pose estimation, compared to feature-point-based SFM algorithms, iii) improve object detections given multiple un-calibrated images, compared to independently detecting objects in single images. Extensive quantitative results on three datasets - LiDAR cars, street-view pedestrians, and Kinect office desktop - verify our theoretical claims.

In this paper we present a framework for the recognition of collective human activities. A collective activity is defined or reinforced by the existence of coherent behavior of individuals in time and space. We call such coherent behavior `Crowd Context'. Examples of collective activities are “queuing in a line” or “talking”. Following, we propose to recognize collective activities using the crowd context and introduce a new scheme for learning it automatically. Our scheme is constructed upon a Random Forest structure which randomly samples variable volume spatio-temporal regions to pick the most discriminating attributes for classification. Unlike previous approaches, our algorithm automatically finds the optimal configuration of spatio-temporal bins, over which to sample the evidence, by randomization. This enables a methodology for modeling crowd context. We employ a 3D Markov Random Field to regularize the classification and localize collective activities in the scene. We demonstrate the flexibility and scalability of the proposed framework in a number of experiments and show that our method outperforms state-of-the art action classification techniques.

In this paper we explore the idea of using high-level semantic concepts, also called attributes, to represent human actions from videos and argue that attributes enable the construction of more descriptive models for human action recognition. We propose a unified framework wherein manually specified attributes are: i) selected in a discriminative fashion so as to account for intra-class variability; ii) coherently integrated with data-driven attributes to make the attribute set more descriptive. Data-driven attributes are automatically inferred from the training data using an information theoretic approach. Our framework is built upon a latent SVM formulation where latent variables capture the degree of importance of each attribute for each action class. We also demonstrate that our attribute-based action representation can be effectively used to design a recognition procedure for classifying novel action classes for which no training samples are available. We test our approach on several publicly available datasets and obtain promising results that quantitatively demonstrate our theoretical claims.

In this paper, we present a novel approach to recognizing human actions from different views by view knowledge transfer. An action is originally modelled as a bag of visual-words (BoVW), which is sensitive to view changes. We argue that, as opposed to visual words, there exist some higher level features which can be shared across views and enable the connection of action models for different views. To discover these features, we use a bipartite graph to model two view-dependent vocabularies, then apply bipartite graph partitioning to co-cluster two vocabularies into visual-word clusters called bilingual-words (i.e., high-level features), which can bridge the semantic gap across view-dependent vocabularies. Consequently, we can transfer a BoVW action model into a bag-of-bilingual-words (BoBW) model, which is more discriminative in the presence of view changes. We tested our approach on the IXMAS data set and obtained very promising results. Moreover, to further fuse view knowledge from multiple views, we apply a Locally Weighted Ensemble scheme to dynamically weight transferred models based on the local distribution structure around each test example. This process can further improve the average recognition rate by about 7%.

The deployment of computer vision algorithms in mobile applications is growing at a rapid pace. A primary component of the computer vision software pipeline is feature extraction, which identifies and encodes relevant image features. We present an embedded heterogeneous multicore design named EFFEX that incorporates novel functional units and memory architecture support, making it capable of increasing mobile vision performance while balancing power and area. We demonstrate this architecture running three common feature extraction algorithms, and show that it is capable of providing significant speedups at low cost. Our simulations show a speedup of as much as 14x for feature extraction with a decrease in energy of 40x for memory accesses.

We propose a novel statistical manifold modeling approach that is capable of classifying poses of object categories from video sequences by simultaneously minimizing the intra-class variability and maximizing inter-pose distance. Following the intuition that an object part based representation and a suitable part selection process may help achieve our purpose, we formulate the part selection problem from a statistical manifold modeling perspective and treat part selection as adjusting the manifold of the object (parameterized by pose) by means of the manifold “alignment” and “expansion” operations. We show that manifold alignment and expansion are equivalent to minimizing the intra-class distance given a pose while increasing the inter-pose distance given an object instance respectively. We formulate and solve this (otherwise intractable) part selection problem as a combinatorial optimization problem using graph analysis techniques. Quantitative and qualitative experimental analysis validates our theoretical claims.

Despite recent successes, pose estimators are still somewhat fragile, and they frequently rely on a precise knowledge of the location of the object. Unfortunately, articulated objects are also very difficult to detect. Knowledge about the articulated nature of these objects, however, can substantially contribute to the task of finding them in an image. It is somewhat surprising, that these two tasks are usually treated entirely separately. In this paper, we propose an Articulated Part-based Model (APM) for jointly detecting objects and estimating their poses. APM recursively represents an object as a collection of parts at multiple levels of detail, from coarse-to-fine, where parts at every level are connected to a coarser level through a parent-child relationship (Fig. 1(b)-Horizontal). Parts are further grouped into part-types (e.g., left-facing head, long stretching arm, etc) so as to model appearance variations (Fig. 1(b)-Vertical). By having the ability to share appearance models of part types and by decomposing complex poses into parent-child pairwise relationships, APM strikes a good balance between model complexity and model richness. Extensive quantitative and qualitative experiment results on public datasets show that APM outperforms state-of-the-art methods. We also show results on PASCAL 2007 - cats and dogs - two highly challenging articulated object categories.

This paper reports a novel algorithm for boot strapping the automatic registration of unstructured 3D point clouds collected using co-registered 3D lidar and omnidirectional camera imagery. Here, we exploit the co-registration of the 3D point cloud with the available camera imagery to associate high dimensional feature descriptors such as scale invariant feature transform (SIFT) or speeded up robust features (SURF) to the 3D points. We first establish putative point correspondence in the high dimensional feature space and then use these correspondences in a random sample consensus (RANSAC) framework to obtain an initial rigid body transformation that aligns the two scans. This initial transformation is then refined in a generalized iterative closest point (ICP) framework. The proposed method is completely data driven and does not require any initial guess on the transformation. We present results from a real world dataset collected by a vehicle equipped with a 3D laser scanner and an omnidirectional camera.

Detecting objects in complex scenes while recovering the scene layout is a critical functionality in many vision-based applications. Inspired by the work of, we advocate the importance of geometric contextual reasoning for object recognition. We start from the intuition that objects' location and pose in the 3D space are not arbitrarily distributed but rather constrained by the fact that objects must lie on one or multiple supporting surfaces. We model such supporting surfaces by means of hidden parameters (i.e. not explicitly observed) and formulate the problem of joint scene reconstruction and object recognition as the one of finding the set of parameters that maximizes the joint probability of having a number of detected objects on K supporting planes given the observations. As a key ingredient for solving this optimization problem, we have demonstrated a novel relationship between object location and pose in the image, and the scene layout parameters (i.e. normal of one or more supporting planes in 3D and camera pose, location and focal length). Using the probabilistic formulation and the above relationship our method has the unique ability to jointly: (i) reduce false alarm and false negative object detection rate; (ii) recover object location and supporting planes within the 3D camera reference system; (iii) infer camera parameters (view point and the focal length) from just one single uncalibrated image. Quantitative and qualitative experimental evaluation on a number of datasets (a novel in-house dataset and label-me on car and pedestrian) demonstrates our theoretical claims.

We address the problem of recognizing the pose of an object category from video sequences capturing the object under small camera movements. This scenario is relevant in applications such as robotic object manipulation or autonomous navigation. We introduce a new algorithm where we model an object category as a collection of non parametric probability densities capturing appearance and geometrical variability within a small area of the viewing sphere for different object instances. By regarding the set of frames of the video as realizations of such probability densities, we cast the problem of object pose classification as the one of matching (i.e., comparing information divergence of) probably density functions in testing and training. Our work can be also related to statistical manifold learning. By performing dimensionality reduction on the manifold of learned PDFs, we show that the embedding in the 3D Euclidean space yield meaningful trajectories which can be parameterized by the pose coordinates on the viewing sphere, this enables an unsupervised learning procedure for pose classification. Our experimental results on both synthesized and real world data show promising results toward the goal of accurate and efficient pose classification of object categories from video sequences.

We propose a novel probabilistic framework for learning visual models of 3D object categories by combining appearance information and geometric constraints. Objects are represented as a coherent ensemble of parts that are consistent under 3D viewpoint transformations. Each part is a collection of salient image features. A generative framework is used for learning a model that captures the relative position of parts within each of the discretized viewpoints. Contrary to most of the existing mixture of viewpoints models, our model establishes explicit correspondences of parts across different viewpoints of the object class. Given a new image, detection and classification are achieved by determining the position and viewpoint of the model that maximize recognition scores of the candidate objects. Our approach is among the first to propose a generative probabilistic framework for 3D object categorization. We test our algorithm on the detection task and the viewpoint classification task by using “car” category from both the Savarese et al. 2007 and PASCAL VOC 2006 datasets. We show promising results in both the detection and viewpoint classification tasks on these two challenging datasets.

In this paper we present a new method for categorizing video sequences capturing different scene classes. This can be seen as a generalization of previous work on scene classification from single images. A scene is represented by a collection of 3D points with an appearance based codeword attached to each point. The cloud of points is recovered by using a robust SFM algorithm applied on the video sequence. A hierarchical structure of histograms located at different locations and at different scales is used to capture the typical spatial distribution of 3D points and codewords in the working volume. The scene is classified by SVM equipped with a histogram matching kernel, similar to. Results on a challenging dataset of 5 scene categories show competitive classification accuracy and superior performance with respect to a state-of-the-art 2D pyramid matching methods applied to individual image frames.

Recognizing object classes and their 3D viewpoints is an important problem in computer vision. Based on a part-based probabilistic representation [31], we propose a new 3D object class model that is capable of recognizing unseen views by pose estimation and synthesis. We achieve this by using a dense, multiview representation of the viewing sphere parameterized by a triangular mesh of viewpoints. Each triangle of viewpoints can be morphed to synthesize new viewpoints. By incorporating 3D geometrical constraints, our model establishes explicit correspondences among object parts across viewpoints. We propose an incremental learning algorithm to train the generative model. A cellphone video clip of an object is first used to initialize model learning. Then the model is updated by a set of unsorted training images without viewpoint labels. We demonstrate the robustness of our model on object detection, viewpoint classification and synthesis tasks. Our model performs superiorly to and on par with state-of-the-art algorithms on the Savarese et al. 2007 and PASCAL datasets in object detection. It outperforms all previous work in viewpoint classification and offers promising results in viewpoint synthesis.

Recognizing and localizing specular (or mirror-like) surfaces from a single image is a great challenge to computer vision. Unlike other materials, the appearance of a specular surface changes as function of the surrounding environment as well as the position of the observer. Even though the reflection on a specular surface has an intrinsic ambiguity that might be resolved by high level reasoning, we argue that we can take advantage of low level features to recognize specular surfaces. This intuition stems from the observation that the surrounding scene is highly distorted when reflected off regions of high curvature or occluding contours. We call these features static specular flows (SSF). We show how to characterize SSF and use them for identifying specular surfaces. To evaluate our result we collect a dataset of 120 images containing specular surfaces. Our algorithm can achieve good performances on this challenging dataset. Particularly, our results outperform other methods that follow a more naive approach.

We propose a novel and robust model to represent and learn generic 3D object categories. We aim to solve the problem of true 3D object categorization for handling arbitrary rotations and scale changes. Our approach is to capture a compact model of an object category by linking together diagnostic parts of the objects from different viewing points. We emphasize on the fact that our "parts" are large and discriminative regions of the objects that are composed of many local invariant features. Instead of recovering a full 3D geometry, we connect these parts through their mutual homographic transformation. The resulting model is a compact summarization of both the appearance and geometry information of the object class. We propose a framework in which learning is done via minimal supervision compared to previous works. Our results on categorization show superior performances to state-of-the-art algorithms such as [23]. Furthermore, we have compiled a new 3D object dataset that consists of 10 different object categories. We have tested our algorithm on this dataset and have obtained highly promising results.

This paper presents a new model of object classes which incorporates appearance and shape information jointly. Modeling objects appearance by distributions of visual words has recently proven successful. Here appearancebased models are augmented by capturing the spatial arrangement of visual words. Compact spatial modeling without loss of discrimination is achieved through the introduction of adaptive vector quantized correlograms, which we call correlatons. Efficiency is further improved by means of integral images. The robustness of our new models to geometric transformations, severe occlusions and missing information is also demonstrated. The accuracy of discrimination of the proposed models is assessed with respect to existing databases with large numbers of object classes viewed under general conditions, and shown to outperform appearance-only models.

We explore the geometry linking the shape of a curved mirror surface to the distortions it produces on a scene it reflects. Our analysis is local and differential. We assume a simple calibrated scene composed of lines passing through a point. We demonstrate that local information about the geometry of the surface may be recovered up to the second order from either the orientation and curvature of the images of two intersecting lines, or from the orientation of the images of three or more intersecting lines. An explicit solution for calculating shape and position of spherical mirror surfaces is given.