Video-language models have shown promise for addressing a range of multimodal tasks for video reasoning, such as video question-answering. However, the inherent computational challenges of processing long video data and increasing model sizes have led to standard approaches that are limited by the number of frames they can process. In this work, we propose the Flexible Frame Selector (FFS), a learnable policy model with a new flexible selection operation, that helps alleviate input context restrictions by enabling video-language models to focus on the most informative frames for the downstream multimodal task, without adding undue processing cost. Our method differentiates from prior work due to its learnability, efficiency, and flexibility. We verify the efficacy of our method on standard video-question answering and reasoning benchmarks, and observe our model can maintain or improve base video-language model accuracy while significantly reducing the number of downstream processed frames.

Video understanding has made huge strides in recent years, relying largely on the power of transformers. As this architecture is notoriously expensive and video data is highly redundant, research into improving efficiency has become particularly relevant. Some creative solutions include token selection and merging. While most methods succeed in reducing the cost of the model and maintaining accuracy, an interesting pattern arises: most methods do not outperform the baseline of randomly discarding tokens. In this paper we take a closer look at this phenomenon and observe 5 principles of the nature of visual tokens. For example, we observe that the value of tokens follows a clear Pareto-distribution where most tokens have remarkably low value, and just a few carry most of the perceptual information. We build on these and further insights to propose a lightweight video model, LITE, that can select a small number of tokens effectively, outperforming state-of-the-art and existing baselines across datasets (Kinetics-400 and Something-Something-V2) in the challenging trade-off of computation (GFLOPs) vs accuracy. Experiments also show that LITE generalizes across datasets and even other tasks without the need for retraining.

Diffusion models have revolutionized generative modeling, enabling unprecedented realism in image and video synthesis. This success has sparked interest in leveraging their representations for visual understanding tasks. While recent works have explored this potential for image generation, the visual understanding capabilities of video diffusion models remain largely uncharted. To address this gap, we systematically compare the same model architecture trained for video versus image generation, analyzing the performance of their latent representations on various downstream tasks including image classification, action recognition, depth estimation, and tracking. Results show that video diffusion models consistently outperform their image counterparts, though we find a striking range in the extent of this superiority. We further analyze features extracted from different layers and with varying noise levels, as well as the effect of model size and training budget on representation and generation quality. This work marks the first direct comparison of video and image diffusion objectives for visual understanding, offering insights into the role of temporal information in representation learning.

We propose a new task and model for dense video object captioning -- detecting, tracking and captioning trajectories of objects in a video. This task unifies spatial and temporal localization in video, whilst also requiring fine-grained visual understanding that is best described by natural language. We propose a unified model, and demonstrate how our end-to-end approach is more accurate and temporally coherent than a multi-stage pipeline combining state-of-the-art detection, tracking, and captioning models. Moreover, we propose a training strategy based on a mixture of disjoint tasks, which allows us to leverage diverse, large-scale datasets which supervise different parts of our model. Although each pretraining task only provides weak supervision, they are complementary and, when combined, result in noteworthy zero-shot ability and serve as strong initialization for additional finetuning to further improve accuracy. We carefully design new metrics capturing all components of our task, and show how we can repurpose existing video grounding datasets (e.g. VidSTG and VLN) for our new task. We show that our model improves upon a number of strong baselines for this new task. Furthermore, we can apply our model to the task of spatial grounding, outperforming prior state-of-the-art on VidSTG and VLN, without explicitly training for it. Our code is available at https://github.com/google-research/scenic.

Despite recent advances in Vision-Language Models (VLMs), long-video understanding remains a challenging problem. Although state-of-the-art long-context VLMs can process around 1000 input frames, they still struggle to effectively leverage this sequence length, and succumb to irrelevant distractors within the context window. We present Dynamic Context Aggregation, an inference strategy for video question-answering that curates the model's input context. We use the VLM itself to iteratively identify and extract the most relevant frames from the video, which are then used for answering. We demonstrate how leveraging more computation at inference-time to select the most relevant context leads to improvements in accuracy, in agreement with recent work on inference-time scaling of LLMs. Moreover, we achieve state-of-the-art results on 4 diverse video question-answering datasets, showing consistent improvements with 3 different VLMs. In particular, our method shines on longer videos which would not otherwise fit in the model's context window: On longer videos of more than 1 hour on LVBench, our approach using a context window of 32K outperforms the same VLM using standard inference with a 700K context window by 2.8 points.

Text-to-image diffusion models excel at translating language prompts into photorealistic images by implicitly grounding textual concepts through their cross-modal attention mechanisms. Recent multi-modal diffusion transformers extend this by introducing joint self-attention over concatenated image and text tokens, enabling richer and more scalable cross-modal alignment. However, a detailed understanding of how and where these attention maps contribute to image generation remains limited. In this paper, we introduce Seg4Diff (Segmentation for Diffusion), a systematic framework for analyzing the attention structures of MM-DiT, with a focus on how specific layers propagate semantic information from text to image. Through comprehensive analysis, we identify a semantic grounding expert layer, a specific MM-DiT block that consistently aligns text tokens with spatially coherent image regions, naturally producing high-quality semantic segmentation masks. We further demonstrate that applying a lightweight fine-tuning scheme with mask-annotated image data enhances the semantic grouping capabilities of these layers and thereby improves both segmentation performance and generated image fidelity. Our findings demonstrate that semantic grouping is an emergent property of diffusion transformers and can be selectively amplified to advance both segmentation and generation performance, paving the way for unified models that bridge visual perception and generation.

The success of the machine learning field has reliably depended on training on large datasets. While effective, this trend comes at an extraordinary cost. This is due to two deeply intertwined factors: the size of models and the size of datasets. While promising research efforts focus on reducing the size of models, the other half of the equation remains fairly mysterious. Indeed, it is surprising that the standard approach to training remains to iterate over and over, uniformly sampling the training dataset. In this paper we explore a series of alternative training paradigms that leverage insights from hard-data-mining and dropout, simple enough to implement and use that can become the new training standard. The proposed Progressive Data Dropout reduces the number of effective epochs to as little as 12.4\% of the baseline. This savings actually do not come at any cost for accuracy. Surprisingly, the proposed method improves accuracy by up to 4.82\%. Our approach requires no changes to model architecture or optimizer, and can be applied across standard training pipelines, thus posing an excellent opportunity for wide adoption. Code can be found here: \url{https://github.com/bazyagami/LearningWithRevision}.

We explore the boundaries of scaling up a multilingual vision and language model, both in terms of size of the components and the breadth of its training task mixture. Our model achieves new levels of performance on a wide-range of varied and complex tasks, including multiple image-based captioning and question-answering tasks, image-based document understanding and few-shot (in-context) learning, as well as object detection, video question answering, and video captioning. Our model advances the state-of-the-art on most vision-and-language benchmarks considered (20+ of them). Finally, we observe emerging capabilities, such as complex counting and multilingual object detection, tasks that are not explicitly in the training mix.

Open-vocabulary semantic segmentation presents the challenge of labeling each pixel within an image based on a wide range of text descriptions. In this work, we introduce a novel cost-based approach to adapt vision-language foundation models, notably CLIP, for the intricate task of semantic segmentation. Through aggregating the cosine similarity score, i. e., the cost volume between image and text embeddings, our method potently adapts CLIP for segmenting seen and unseen classes by fine-tuning its encoders, addressing the challenges faced by existing methods in handling unseen classes. Building upon this, we explore methods to effectively aggregate the cost volume considering its multi-modal nature of being established between image and text embeddings. Furthermore, we examine various methods for efficiently fine-tuning CLIP.

The most performant spatio-temporal action localisation models use external person proposals and complex external memory banks. We propose a fully end-to-end, purely-transformer based model that directly ingests an input video, and outputs tubelets - a sequence of bounding boxes and the action classes at each frame. Our flexible model can be trained with either sparse bounding-box supervision on individual frames, or full tubelet annotations. And in both cases, it predicts coherent tubelets as the output. Moreover, our end-to-end model requires no additional pre-processing in the form of proposals, or post-processing in terms of non-maximal suppression. We perform extensive ablation experiments, and significantly advance the state-of-the-art on five different spatio-temporal action localisation benchmarks with both sparse keyframes and full tubelet annotations.

Vision-Language models (VLMs) have excelled in the image-domain- especially in zero-shot settings- thanks to the availability of vast pretraining data (i.e., paired image-text samples). However for videos, such paired data is not as abundant. Therefore, video- VLMs are usually designed by adapting pretrained image- VLMs to the video-domain, instead of training from scratch. All such recipes rely on aug-menting visual embeddings with temporal information (i.e., image -+ video), often keeping text embeddings unchanged or even being discarded. In this paper, we argue the contrary, that better video- VLMs can be designed by focusing more on augmenting text, rather than visual information. More specifically, we introduce Video-conditioned Text Representations (Vi c TR): a form of text embeddings optimized w.r.t. vi-sual embeddings, creating a more-flexible contrastive latent space. Our model canfurther make use offreely-available semantic information, in the form of visually- grounded aux-iliary text (e.g. object or scene information). We evaluate our model on few-shot, zero-shot (HMDB-51, UCF-10l), short-form (Kinetics-400) and long-form (Charades) activ-ity recognition benchmarks, showing strong performance among video-VLMs.

As foundation models become more popular, there is a growing need to efficiently finetune them for downstream tasks. Although numerous adaptation methods have been proposed, they are designed to be efficient only in terms of how many parameters are trained. They, however, typi-cally still require backpropagating gradients throughout the model, meaning that their training-time and -memory cost does not reduce as significantly. We propose an adaptation method which does not back-propagate gradients through the backbone. We achieve this by designing a lightweight network in parallel that oper-ates on features from the frozen, pretrained backbone. As a result, our method is efficient not only in terms of parame-ters, but also in training-time and memory usage. Our approach achieves state-of-the-art accuracy-parameter trade-offs on the popular VTAB benchmark, and we further show how we outperform prior works with respect to training-time and -memory usage too. We further demonstrate the training efficiency and scalability of our method by adapting a vision transformer backbone of 4 billion parameters for the computationally demanding task of video classifi-cation, without any intricate model parallelism. Here, we outperform a prior adaptor-based method which could only scale to a 1 billion parameter backbone, or fully-finetuning a smaller backbone, with the same GPU and less training time. To facilitate further research, we release code at https://github.com/google-researchlscenic.

Large language models have achieved great success in recent years, so as their variants in vision. Existing vision-language models can describe images in natural languages, answer visual-related questions, or perform complex reasoning about the image. However, it is yet unclear how localization tasks, such as word grounding or referring localization, can be performed using large language models. In this work, we aim to develop a vision-language model that can take locations, for example, a set of points or boxes, as either inputs or outputs. When taking locations as inputs, the model performs location-conditioned captioning, which generates captions for the indicated object or region. When generating locations as outputs, our model regresses pixel coordinates for each output word generated by the language model, and thus performs dense word grounding. Our model is pre-trained on the Localized Narrative dataset, which contains pixel-word-aligned captioning from human attention. We show our model can be applied to various location-aware vision-language tasks, including referring localization, location-conditioned captioning, and dense object captioning, archiving state-of-the-art performance on RefCOCO and Visual Genome.

An ideal model for dense video captioning - predicting captions localized temporally in a video - should be able to handle long input videos, predict rich, detailed textual descriptions, and be able to produce outputs before processing the entire video. Current state-of-the-art models, however, process a fixed number of downsampled frames, and make a single full prediction after seeing the whole video. We propose a streaming dense video captioning model that consists of two novel components: First, we propose a new memory module, based on clustering incoming tokens, which can handle arbitrarily long videos as the memory is of a fixed size. Second, we develop a streaming decoding algorithm that enables our model to make predictions before the entire video has been processed. Our model achieves this streaming ability, and significantly improves the state-of-the-art on three dense video captioning benchmarks: ActivityNet, YouCook2 and ViTT. Our code is released at https://github.com/google-research/scenic.

Hammad Ayyubi, Tianqi Liu, Arsha Nagrani, Xudong Lin, Mingda Zhang, Anurag Arnab, Feng Han, Yukun Zhu, Xuande Feng, Kevin Zhang, Jialu Liu, Shih-Fu Chang. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Protecting and restoring forest ecosystems is critical for bio-diversity conservation and climate change mitigation. Forest monitoring on a global scale is essential for prioritizing and assessing conservation efforts. Satellite-based remote sensing is the only viable solution for providing global coverage, but to date, large-scale forest monitoring is limited mostly to single modalities or single time points. In this paper, we present a dataset consisting of data from five public satellites for recognizing forest plantations and planted tree species across the globe. Each satellite modality consists of a multi-year time series. The dataset, named "Planted" includes over 2M examples of 64 tree label classes (46 genera and 40 distinctive species), distributed among 41 countries. This dataset is released to foster research in forest monitoring using multimodal, multi-scale, multi-temporal data sources. Additionally, we present initial baseline results and evaluate modality fusion and data augmentation approaches for this dataset.

The visual medium (images and videos) naturally contains a large amount of information redundancy, thereby providing a great opportunity for leveraging efficiency in processing. While Vision Transformer (ViT) based models scale effectively to large data regimes, they fail to capitalize on this inherent redundancy, leading to higher computational costs. Mixture of Experts (MoE) networks demonstrate scalability while maintaining same inference-time costs, but they come with a larger parameter footprint. We present Mixture of Nested Experts (MoNE), which utilizes a nested structure for experts, wherein individual experts fall on an increasing compute-accuracy curve. Given a compute budget, MoNE learns to dynamically choose tokens in a priority order, and thus redundant tokens are processed through cheaper nested experts. Using this framework, we achieve equivalent performance as the baseline models, while reducing inference time compute by over two-fold. We validate our approach on standard image and video datasets - ImageNet-21K, Kinetics400, and Something-Something-v2. We further highlight MoNE's adaptability by showcasing its ability to maintain strong performance across different inference-time compute budgets on videos, using only a single trained model.

Large-scale vision-language models like CLIP have demonstrated impressive open-vocabulary capabilities for image-level tasks, excelling in recognizing what objects are present. However, they struggle with pixel-level recognition tasks like semantic segmentation, which require understanding where the objects are located. In this work, we propose a novel method, PixelCLIP, to adapt the CLIP image encoder for pixel-level understanding by guiding the model on where, which is achieved using unlabeled images and masks generated from vision foundation models such as SAM and DINO. To address the challenges of leveraging masks without semantic labels, we devise an online clustering algorithm using learnable class names to acquire general semantic concepts. PixelCLIP shows significant performance improvements over CLIP and competitive results compared to caption-supervised methods in open-vocabulary semantic segmentation.

We propose Token Turing Machines (TTM), a sequential, autoregressive Transformer model with memory for real-world sequential visual understanding. Our model is inspired by the seminal Neural Turing Machine, and has an external memory consisting of a set of tokens which summarise the previous history (i.e., frames). This memory is efficiently addressed, read and written using a Transformer as the processing unit/controller at each step. The model's memory module ensures that a new observation will only be processed with the contents of the memory (and not the entire history), meaning that it can efficiently process long sequences with a bounded computational cost at each step. We show that TTM outperforms other alternatives, such as other Transformer models designed for long sequences and recurrent neural networks, on two real-world sequential visual understanding tasks: online temporal activity detection from videos and vision-based robot action policy learning. Code is publicly available at: https://github.com/google-research/scenic/tree/main/scenic/projects/token.turing.

Current state-of-the-art video models process a video clip as a long sequence of spatio-temporal tokens. However, they do not explicitly model objects, their interactions across the video, and instead process all the tokens in the video. In this paper, we investigate how we can use knowledge of objects to design better video models, namely to process fewer tokens and to improve recognition accuracy. This is in contrast to prior works which either drop tokens at the cost of accuracy, or increase accuracy whilst also increasing the computation required. First, we propose an object-guided token sampling strategy that enables us to retain a small fraction of the input tokens with minimal impact on accuracy. And second, we propose an object-aware attention module that enriches our feature representation with object information and improves overall accuracy. Our resulting model, ObjectViViT, achieves better performance when using fewer tokens than strong baselines. In particular, we match our baseline with 30%, 40%, and 60% of the input tokens on SomethingElse, Something-something v2, and Epic-Kitchens, respectively. When we use Object-ViViT to process the same number of tokens as our baseline, we improve by 0.6 to 4.2 points on these datasets.

Can we leverage the audiovisual information already present in video to improve self-supervised representation learning? To answer this question, we study various pre-training architectures and objectives within the masked autoencoding framework, motivated by the success of similar methods in natural language and image understanding. We show that we can achieve significant improvements on audiovisual downstream classification tasks, surpassing the state-of-the-art on VGGSound and AudioSet. Furthermore, we can leverage our audiovisual pretraining scheme for multiple unimodal downstream tasks using a single audiovisual pretrained model. We additionally demonstrate the transferability of our representations, achieving state-of-the-art audiovisual results on Epic Kitchens without pre-training specifically for this dataset.

While large-scale image-text pretrained models such as CLIP have been used for multiple video-level tasks on trimmed videos, their use for temporal localization in untrimmed videos is still a relatively unexplored task. We design a new approach for this called UnLoc, which uses pretrained image and text towers, and feeds tokens to a video-text fusion model. The output of the fusion module are then used to construct a feature pyramid in which each level connects to a head to predict a per-frame relevancy score and start/end time displacements. Unlike previous works, our architecture enables Moment Retrieval, Temporal Localization, and Action Segmentation with a single stage model, without the need for action proposals, motion based pretrained features or representation masking. Unlike specialized models, we achieve state of the art results on all three different localization tasks with a unified approach. Code is available at: https://github.com/google-research/scenic.

The scaling of Transformers has driven breakthrough capabilities for language models. At present, the largest large language models (LLMs) contain upwards of 100B parameters. Vision Transformers (ViT) have introduced the same architecture to image and video modelling, but these have not yet been successfully scaled to nearly the same degree; the largest dense ViT contains 4B parameters (Chen et al., 2022). We present a recipe for highly efficient and stable training of a 22B-parameter ViT (ViT-22B) and perform a wide variety of experiments on the resulting model. When evaluated on downstream tasks (often with a lightweight linear model on frozen features), ViT-22B demonstrates increasing performance with scale. We further observe other interesting benefits of scale, including an improved tradeoff between fairness and performance, state-of-the-art alignment to human visual perception in terms of shape/texture bias, and improved robustness. ViT-22B demonstrates the potential for "LLM-like" scaling in vision, and provides key steps towards getting there.

Humans have the ability to adapt the type of information they use, the procedure they employ, and the amount of time they spend when solving problems. However, most standard neural networks have a fixed function type and computation budget regardless of the sample’s nature or difficulty. Adaptivity is a powerful paradigm as it not only imbues practitioners with flexibility pertaining to the downstream usage of these models but can also serve as a powerful inductive bias for solving certain challenging classes of problems. In this work, we introduce a new approach called AdaTape, which allows for dynamic computation in neural networks through adaptive tape tokens. AdaTape utilizes an elastic input sequence by equipping an architecture with a dynamic read-and-write tape. Specifically, we adaptively generate input sequences using tape tokens obtained from a tape bank which can be either trainable or derived from input data. We examine the challenges and requirements to obtain dynamic sequence content and length, and propose the Adaptive Tape Reading (ATR) algorithm to achieve both goals. Through extensive experiments on image recognition tasks, we show that AdaTape can achieve better performance while maintaining the computational cost. To facilitate further research, we have released code at https://github.com/google-research/scenic/tree/main/scenic/projects/adatape.

We aim to investigate whether end-to-end learning of visual reasoning can be achieved with general-purpose neural networks, with the help of visual pretraining. A positive result would refute the common belief that explicit visual abstraction (e.g. object detection) is essential for compositional generalization on visual reasoning, and confirm the feasibility of a neural network ''generalist'' to solve visual recognition and reasoning tasks. We propose a simple and general self-supervised framework which ''compresses'' each video frame into a small set of tokens with a transformer network, and reconstructs the remaining frames based on the compressed temporal context. To minimize the reconstruction loss, the network must learn a compact representation for each image, as well as capture temporal dynamics and object permanence from temporal context. We perform evaluation on two visual reasoning benchmarks, CATER and ACRE. We observe that pretraining is essential to achieve compositional generalization for end-to-end visual reasoning. Our proposed framework outperforms traditional supervised pretraining, including image classification and explicit object detection, by large margins.

Modelling long-range dependencies is critical for scene understanding tasks in computer vision. Although convolution neural networks (CNNs) have excelled in many vision tasks, they are still limited in capturing long-range structured relationships as they typically consist of layers of local kernels. A fully-connected graph, such as the self-attention operation in Transformers, is beneficial for such modelling, however, its computational overhead is prohibitive. In this paper, we propose a dynamic graph message passing network, that significantly reduces the computational complexity compared to related works modelling a fully-connected graph. This is achieved by adaptively sampling nodes in the graph, conditioned on the input, for message passing. Based on the sampled nodes, we dynamically predict node-dependent filter weights and the affinity matrix for propagating information between them. This formulation allows us to design a self-attention module, and more importantly a new Transformer-based backbone network, that we use for both image classification pretraining, and for addressing various downstream tasks (e.g. object detection, instance and semantic segmentation). Using this model, we show significant improvements with respect to strong, state-of-the-art baselines on four different tasks. Our approach also outperforms fully-connected graphs while using substantially fewer floating-point operations and parameters. Code and models will be made publicly available at https://github.com/fudan-zvg/DGMN2.

Scenic is an open-source11https://github.com/google-research/scenic JAX library with a focus on transformer-based models for computer vision research and beyond. The goal of this toolkit is to facilitate rapid experimentation, prototyping, and research of new architectures and models. Scenic supports a diverse range of tasks (e.g., classification, segmentation, detection) and facilitates working on multi-modal problems, along with GPU/TPU support for large-scale, multi-host and multi-device training. Scenic also offers optimized implementations of state-of-the-art research models spanning a wide range of modalities. Scenic has been successfully used for numerous projects and published papers and continues serving as the library of choice for rapid prototyping and publication of new research ideas.

Recent advances in deep learning have relied on large, labelled datasets to train high-capacity models. However, collecting large datasets in a time- and cost-efficient manner often results in label noise. We present a method for learning from noisy labels that leverages similarities between training examples in feature space, encouraging the prediction of each example to be similar to its nearest neighbours. Compared to training algorithms that use multiple models or distinct stages, our approach takes the form of a simple, additional regularization term. It can be interpreted as an inductive version of the classical, transductive label propagation algorithm. We thoroughly evaluate our method on datasets evaluating both synthetic (CIFAR-10, CIFAR-100) and realistic (mini-WebVision, WebVision, Clothing1M, mini-ImageNet-Red) noise, and achieve competitive or state-of-the-art accuracies across all of them.

Recent video and language pretraining frameworks lack the ability to generate sentences. We present Multimodal Video Generative Pretraining (MV-GPT), a new pretraining framework for learning from unlabelled videos which can be effectively used for generative tasks such as multimodal video captioning. Unlike recent video-language pretraining frameworks, our framework trains both a multimodal video encoder and a sentence decoder jointly. To overcome the lack of captions in unlabelled videos, we leverage the future utterance as an additional text source and propose a bidirectional generation objective - we generate future utterances given the present mulitmodal context, and also the present utterance given future observations. With this objective, we train an encoder-decoder model end-to-end to generate a caption from raw pixels and transcribed speech directly. Our model achieves state-of the-art performance for multimodal video captioning on four standard benchmarks, as well as for other video understanding tasks such as VideoQA, video retrieval and action classification.

Video understanding requires reasoning at multiple spatiotemporal resolutions – from short fine-grained motions to events taking place over longer durations. Although transformer architectures have recently advanced the state-of-the-art, they have not explicitly modelled different spatiotemporal resolutions. To this end, we present Multiview Transformers for Video Recognition (MTV). Our model consists of separate encoders to represent different views of the input video with lateral connections to fuse information across views. We present thorough ablation studies of our model and show that MTV consistently performs better than single-view counterparts in terms of accuracy and computational cost across a range of model sizes. Furthermore, we achieve state-of-the-art results on six standard datasets, and improve even further with large-scale pretraining. Code and checkpoints are available at: https://github.com/google-research/scenic.

We present pure-transformer based models for video classification, drawing upon the recent success of such models in image classification. Our model extracts spatiotemporal tokens from the input video, which are then encoded by a series of transformer layers. In order to handle the long sequences of tokens encountered in video, we propose several, efficient variants of our model which factorise the spatial- and temporal-dimensions of the input. Although transformer-based models are known to only be effective when large training datasets are available, we show how we can effectively regularise the model during training and leverage pretrained image models to be able to train on comparatively small datasets. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple video classification benchmarks including Kinetics 400 and 600, Epic Kitchens, Something-Something v2 and Moments in Time, outperforming prior methods based on deep 3D convolutional networks.

Accurate video understanding involves reasoning about the relationships between actors, objects and their environment, often over long temporal intervals. In this paper, we propose a message passing graph neural network that explicitly models these spatio-temporal relations and can use explicit representations of objects, when supervision is available, and implicit representations otherwise. Our formulation generalises previous structured models for video understanding, and allows us to study how different design choices in graph structure and representation affect the model’s performance. We demonstrate our method on two different tasks requiring relational reasoning in videos – spatio-temporal action detection on AVA and UCF101-24, and video scene graph classification on the recent Action Genome dataset – and achieve state-of-the-art results on all three datasets. Furthermore, we show quantitatively and qualitatively how our method is able to more effectively model relationships between relevant entities in the scene.

Learning effective visual representations that generalize well without human supervision is a fundamental problem in order to apply Machine Learning to a wide variety of tasks. Recently, two families of self-supervised methods, contrastive learning and latent bootstrapping, exemplified by SimCLR and BYOL respectively, have made significant progress. In this work, we hypothesize that adding explicit information compression to these algorithms yields better and more robust representations. We verify this by developing SimCLR and BYOL formulations compatible with the Conditional Entropy Bottleneck (CEB) objective, allowing us to both measure and control the amount of compression in the learned representation, and observe their impact on downstream tasks. Furthermore, we explore the relationship between Lipschitz continuity and compression, showing a tractable lower bound on the Lipschitz constant of the encoders we learn. As Lipschitz continuity is closely related to robustness, this provides a new explanation for why compressed models are more robust. Our experiments confirm that adding compression to SimCLR and BYOL significantly improves linear evaluation accuracies and model robustness across a wide range of domain shifts. In particular, the compressed version of BYOL achieves 76.0% Top-1 linear evaluation accuracy on ImageNet with ResNet-50, and 78.8% with ResNet-50 2x.

Humans perceive the world by concurrently processing and fusing high-dimensional inputs from multiple modalities such as vision and audio. Machine perception models, in stark contrast, are typically modality-specific and optimised for unimodal benchmarks.A common approach for building multimodal models is to simply combine multiple of these modality-specific architectures using late-stage fusion of final representations or predictions ('late-fusion').Instead, we introduce a novel transformer based architecture that uses 'attention bottlenecks' for modality fusion at multiple layers. Compared to traditional pairwise self-attention, these bottlenecks force information between different modalities to pass through a small number of '`bottleneck' latent units, requiring the model to collate and condense the most relevant information in each modality and only share what is necessary. We find that such a strategy improves fusion performance, at the same time reducing computational cost. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple audio-visual classification benchmarks including Audioset, Epic-Kitchens and VGGSound. All code and models will be released.

In this paper, we introduce a novel visual representation learning which relies on a handful of adaptively learned tokens, and which is applicable to both image and video understanding tasks. Instead of relying on hand-designed splitting strategies to obtain visual tokens and processing a large number of densely sampled patches for attention, our approach learns to mine important tokens in visual data. This results in efficiently and effectively finding a few important visual tokens and enables modeling of pairwise attention between such tokens, over a longer temporal horizon for videos, or the spatial content in image frames. Our experiments demonstrate strong performance on several challenging benchmarks for video recognition tasks. Importantly, due to our tokens being adaptive, we accomplish competitive results at significantly reduced computational cost. We establish new state-of-the-arts on multiple video datasets, including Kinetics-400, Kinetics-600, Charades, and AViD.

Modelling long-range dependencies is critical for scene understanding tasks in computer vision. Although CNNs have excelled in many vision tasks, they are still limited in capturing long-range structured relationships as they typically consist of layers of local kernels. A fully-connected graph is beneficial for such modelling, however, its computational overhead is prohibitive. We propose a dynamic graph message passing network, that significantly reduces the computational complexity compared to related works modelling a fully-connected graph. This is achieved by adaptively sampling nodes in the graph, conditioned on the input, for message passing. Based on the sampled nodes, we dynamically predict node-dependent filter weights and the affinity matrix for propagating information between them. Using this model, we show significant improvements with respect to strong, state-of-the-art baselines on three different tasks and backbone architectures. Our approach also outperforms fully-connected graphs while using substantially fewer floating-point operations and parameters.

Personal robots and driverless cars need to be able to operate in novel environments and thus quickly and efficiently learn to recognise new object classes. We address this problem by considering the task of video object segmentation. Previous accurate methods for this task finetune a model using the first annotated frame, and/or use additional inputs such as optical flow and complex post-processing. In contrast, we develop a fast, causal algorithm that requires no finetuning, auxiliary inputs or post-processing, and segments a variable number of objects in a single forward-pass. We represent an object with clusters, or "visual words", in the embedding space, which correspond to object parts in the image space. This allows us to robustly match to the reference objects throughout the video, because although the global appearance of an object changes as it undergoes occlusions and deformations, the appearance of more local parts may stay consistent. We learn these visual words in an unsupervised manner, using meta-learning to ensure that our training objective matches our inference procedure. We achieve comparable accuracy to finetuning based methods (whilst being 1 to 2 orders of magnitude faster), and state-of-the-art in terms of speed/accuracy trade-offs on four video segmentation datasets. Code is available at https://github.com/harkiratbehl/MetaVOS.

We present a bundle-adjustment-based algorithm for recovering accurate 3D human pose and meshes from monocular videos. Unlike previous algorithms which operate on single frames, we show that reconstructing a person over an entire sequence gives extra constraints that can resolve ambiguities. This is because videos often give multiple views of a person, yet the overall body shape does not change and 3D positions vary slowly. Our method improves not only on standard mocap-based datasets like Human 3.6M -- where we show quantitative improvements -- but also on challenging in-the-wild datasets such as Kinetics. Building upon our algorithm, we present a new dataset of more than 3 million frames of YouTube videos from Kinetics with automatically generated 3D poses and meshes. We show that retraining a single-frame 3D pose estimator on this data improves accuracy on both real-world and mocap data by evaluating on the 3DPW and HumanEVA datasets.

We propose a 2D multi-level appearance representation of the human body in RGB images, spatially modelled using a fully-connected graphical model. The appearance model is based on a CNN body part detector, which uses shared features in a cascade architecture to simultaneously detect body parts with different levels of granularity. We use a fully-connected Conditional Random Field (CRF) as our spatial model, over which approximate inference is efficiently performed using the Mean-Field algorithm, implemented as a Recurrent Neural Network (RNN). The stronger visual support from body parts with different levels of granularity, along with the fully-connected pairwise spatial relations, which have their weights learnt by the model, improve the performance of the bottom-up part detector. We adopt an end-to-end training strategy to leverage the potential of both our appearance and spatial models, and achieve competitive results on the MPII and LSP datasets.

Deep Neural Networks (DNNs) have demonstrated exceptional performance on most recognition tasks such as image classification and segmentation. However, they have also been shown to be vulnerable to adversarial examples. This phenomenon has recently attracted a lot of attention but it has not been extensively studied on multiple, large-scale datasets and structured prediction tasks such as semantic segmentation which often require more specialised networks with additional components such as CRFs, dilated convolutions, skip-connections and multiscale processing. In this paper, we present what to our knowledge is the first rigorous evaluation of adversarial attacks on modern semantic segmentation models, using two large-scale datasets. We analyse the effect of different network architectures, model capacity and multiscale processing, and show that many observations made on the task of classification do not always transfer to this more complex task. Furthermore, we show how mean-field inference in deep structured models, multiscale processing (and more generally, input transformations) naturally implement recently proposed adversarial defenses. Our observations will aid future efforts in understanding and defending against adversarial examples. Moreover, in the shorter term, we show how to effectively benchmark robustness and show which segmentation models should currently be preferred in safety-critical applications due to their inherent robustness.

Semantic segmentation and other pixel-level labeling tasks have made significant progress recently due to the deep learning paradigm. Many state-of-the-art structured prediction methods also include a random field model with a hand-crafted Gaussian potential to model spatial priors and label consistencies and feature-based image conditioning. These random field models with image conditioning typically require computationally demanding filtering techniques during inference. In this paper, we present a new inference and learning framework which can learn arbitrary pairwise conditional random field (CRF) potentials. Both standard spatial and high-dimensional bilateral kernels are considered. In addition, we introduce a new type of potential function which is image-dependent like the bilateral kernel, but an order of magnitude faster to compute since only spatial convolutions are employed. It is empirically demonstrated that such learned potentials can improve segmentation accuracy and that certain label-class interactions are indeed better modeled by a non-Gaussian potential. Our framework is evaluated on several public benchmarks for semantic segmentation with improved performance compared to previous state-of-the-art CNN+CRF models.

Semantic segmentation and object detection research have recently achieved rapid progress. However, the former task has no notion of different instances of the same object, and the latter operates at a coarse, bounding-box level. We propose an Instance Segmentation system that produces a segmentation map where each pixel is assigned an object class and instance identity label. Most approaches adapt object detectors to produce segments instead of boxes. In contrast, our method is based on an initial semantic segmentation module, which feeds into an instance subnetwork. This subnetwork uses the initial category-level segmentation, along with cues from the output of an object detector, within an end-to-end CRF to predict instances. This part of our model is dynamically instantiated to produce a variable number of instances per image. Our end-to-end approach requires no post-processing and considers the image holistically, instead of processing independent proposals. Therefore, unlike some related work, a pixel cannot belong to multiple instances. Furthermore, far more precise segmentations are achieved, as shown by our substantial improvements at high APr thresholds.

It is not always possible to recognise objects and infer material properties for a scene from visual cues alone, since objects can look visually similar whilst being made of very different materials. In this paper, we therefore present an approach that augments the available dense visual cues with sparse auditory cues in order to estimate dense object and material labels. Since estimates of object class and material properties are mutually informative, we optimise our multi-output labelling jointly using a random-field framework. We evaluate our system on a new dataset with paired visual and auditory data that we make publicly available. We demonstrate that this joint estimation of object and material labels significantly outperforms the estimation of either category in isolation.