We study self-supervised procedure learning, which discovers key steps and their order from a set of unlabeled videos. Previous methods typically learn frame-to-frame correspondences between videos before determining key steps and their order. However, their performance often suffers from order variations, background/redundant frames, and repeated actions. To overcome these challenges, we propose a self-supervised framework, which utilizes a fused Gromov-Wasserstein optimal transport with a structural prior for frame-to-frame mapping. However, optimizing only for the above temporal alignment may lead to degenerate solutions, where all frames are mapped to a small cluster in the embedding space and thus every video is assigned to just one key step. To address that issue, we integrate a contrastive regularization, which maps different frames to various points, avoiding trivial solutions. Finally, extensive experiments on egocentric and third-person benchmarks demonstrate our superior performance over prior works, including OPEL which relies on a classical Kantorovich optimal transport with an optimality prior.

We introduce a novel approach for simultaneous self-supervised video alignment and action segmentation based on a unified optimal transport framework. In particular, we first tackle self-supervised video alignment by developing a fused Gromov-Wasserstein optimal transport formulation with a structural prior, which trains efficiently on GPUs and needs only a few iterations for solving the optimal transport problem. Our single-task method achieves the state-of-the-art performance on multiple video alignment benchmarks and outperforms VAVA, which relies on a traditional Kantorovich optimal transport formulation with an optimality prior. Furthermore, we extend our approach by proposing a unified optimal transport framework for joint self-supervised video alignment and action segmentation, which requires training and storing a single model and saves both time and memory consumption as compared to two different single-task models. Extensive evaluations on several video alignment and action segmentation datasets demonstrate that our multi-task method achieves comparable video alignment yet superior action segmentation results over previous methods in video alignment and action segmentation respectively. Finally, to the best of our knowledge, this is the first work to unify video alignment and action segmentation into a single model. Our code is available on our research website: https://retrocausal.ai/research/.

This paper presents a 2D skeleton-based action segmentation method with applications in fine-grained human activity recognition. In contrast with state-of-the-art methods which directly take sequences of 3D skeleton coordinates as inputs and apply Graph Convolutional Networks (GCNs) for spatiotemporal feature learning, our main idea is to use sequences of 2D skeleton heatmaps as inputs and employ Temporal Convolutional Networks (TCNs) to extract spatiotemporal features. Despite lacking 3D information, our approach yields comparable/superior performances and better robustness against missing keypoints than previous methods on action segmentation datasets. Moreover, we improve the performances further by using both 2D skeleton heatmaps and RGB videos as inputs. To our best knowledge, this is the first work to utilize 2D skeleton heatmap inputs and the first work to explore 2D skeleton+RGB fusion for action segmentation.

This paper presents an unsupervised transformer-based framework for temporal activity segmentation which leverages not only frame-level cues but also segment-level cues. This is in contrast with previous methods which often rely on frame-level information only. Our approach begins with a frame-level prediction module which estimates framewise action classes via a transformer encoder. The frame-level prediction module is trained in an unsupervised manner via temporal optimal transport. To exploit segment-level information, we utilize a segment-level prediction module and a frame-to-segment alignment module. The former includes a transformer decoder for estimating video transcripts, while the latter matches frame-level features with segmentlevel features, yielding permutation-aware segmentation results. Moreover, inspired by temporal optimal transport, we introduce simple-yet-effective pseudo labels for unsupervised training of the above modules. Our experiments on four public datasets, i.e., 50 Salads, YouTube Instructions, Breakfast, and Desktop Assembly show that our approach achieves comparable or better performance than previous methods in unsupervised activity segmentation.

The lack of large-scale real datasets with annotations makes transfer learning a necessity for video activity understanding. We aim to develop an effective method for few-shot transfer learning for first-person action classification. We leverage independently trained local visual cues to learn representations that can be transferred from a source domain, which provides primitive action labels, to a different target domain using only a handful of examples. Visual cues we employ include object-object interactions, hand grasps and motion within regions that are a function of hand locations. We employ a framework based on meta-learning to extract the distinctive and domain invariant components of the deployed visual cues. This enables transfer of action classification models across public datasets captured with diverse scene and action configurations. We present comparative results of our transfer learning methodology and report superior results over state-of-the-art action classification approaches for both inter-class and inter-dataset transfer.

We present a novel approach for unsupervised activity segmentation which uses video frame clustering as a pretext task and simultaneously performs representation learning and online clustering. This is in contrast with prior works where representation learning and clustering are often performed sequentially. We leverage temporal information in videos by employing temporal optimal transport. In particular, we incorporate a temporal regularization term which preserves the temporal order of the activity into the standard optimal transport module for computing pseudo-label cluster assignments. The temporal optimal transport module enables our approach to learn effective representations for unsupervised activity segmentation. Furthermore, previous methods require storing learned features for the entire dataset before clustering them in an offline manner, whereas our approach processes one mini-batch at a time in an online manner. Extensive evaluations on three public datasets, i.e. 50-Salads, YouTube Instructions, and Breakfast, and our dataset, i.e., Desktop Assembly, show that our approach performs on par with or better than previous methods, despite having significantly less memory constraints.

We introduce a novel approach for temporal activity segmentation with timestamp supervision. Our main contribution is a graph convolutional network, which is learned in an end-to-end manner to exploit both frame features and connections between neighboring frames to generate dense framewise labels from sparse timestamp labels. The gener-ated dense framewise labels can then be used to train the segmentation model. In addition, we propose a framework for alternating learning of both the segmentation model and the graph convolutional model, which first initializes and then iteratively refines the learned models. Detailed experiments on four public datasets, including 50 Salads, GTEA, Breakfast, and Desktop Assembly, show that our method is superior to the multi-layer perceptron baseline, while performing on par with or better than the state of the art in temporal activity segmentation with timestamp supervision.

We present a self-supervised approach for learning video representations using temporal video alignment as a pretext task, while exploiting both frame-level and video-level information. We leverage a novel combination of temporal alignment loss and temporal regularization terms, which can be used as supervision signals for training an encoder network. Specifically, the temporal alignment loss (i.e., Soft-DTW) aims for the minimum cost for temporally aligning videos in the embedding space. However, optimizing solely for this term leads to trivial solutions, particularly, one where all frames get mapped to a small cluster in the embedding space. To overcome this problem, we propose a temporal regularization term (i.e., Contrastive-IDM) which encourages different frames to be mapped to different points in the embedding space. Extensive evaluations on various tasks, including action phase classification, action phase progression, and fine-grained frame retrieval, on three datasets, namely Pouring, Penn Action, and IKEA ASM, show superior performance of our approach over state-of-the-art methods for self-supervised representation learning from videos. In addition, our method provides significant performance gain where labeled data is lacking.

Inexpensive sensing and computation, as well as insurance innovations, have made smart dashboard cameras ubiquitous. Increasingly, simple model-driven computer vision algorithms focused on lane departures or safe following distances are finding their way into these devices. Unfortunately, the long-tailed distribution of road hazards means that these hand-crafted pipelines are inadequate for driver safety systems. We propose to apply data-driven anomaly detection ideas from deep learning to dashcam videos, which hold the promise of bridging this gap. Unfortunately, there exists almost no literature applying anomaly understanding to moving cameras, and correspondingly there is also a lack of relevant datasets. To counter this issue, we present a large and diverse dataset of truck dashcam videos, namely RetroTrucks, that includes normal and anomalous driving scenes. We apply: (i) one-class classification loss and (ii) reconstruction-based loss, for anomaly detection on RetroTrucks as well as on existing static-camera datasets. We introduce formulations for modeling object interactions in this context as priors. Our experiments indicate that our dataset is indeed more challenging than standard anomaly detection datasets, and previous anomaly detection methods do not perform well here out-of-the-box. In addition, we share insights into the behavior of these two important families of anomaly detection approaches on dashcam data.

Recent data-driven approaches to scene interpretation predominantly pose inference as an end-to-end black-box mapping, commonly performed by a Convolutional Neural Network (CNN). However, decades of work on perceptual organization in both human and machine vision suggest that there are often intermediate representations that are intrinsic to an inference task, and which provide essential structure to improve generalization. In this work, we explore an approach for injecting prior domain structure into neural network training by supervising hidden layers of a CNN with intermediate concepts that normally are not observed in practice. We formulate a probabilistic framework which formalizes these notions and predicts improved generalization via this deep supervision method. One advantage of this approach is that we are able to train only from synthetic CAD renderings of cluttered scenes, where concept values can be extracted, but apply the results to real images. Our implementation achieves the state-of-the-art performance of 2D/3D keypoint localization and image classification on real image benchmarks including KITTI, PASCAL VOC, PASCAL3D+, IKEA, and CIFAR100. We provide additional evidence that our approach outperforms alternative forms of supervision, such as multi-task networks.

Monocular 3D object parsing is highly desirable in various scenarios including occlusion reasoning and holistic scene interpretation. We present a deep convolutional neural network (CNN) architecture to localize semantic parts in 2D image and 3D space while inferring their visibility states, given a single RGB image. Our key insight is to exploit domain knowledge to regularize the network by deeply supervising its hidden layers, in order to sequentially infer intermediate concepts associated with the final task. To acquire training data in desired quantities with ground truth 3D shape and relevant concepts, we render 3D object CAD models to generate large-scale synthetic data and simulate challenging occlusion configurations between objects. We train the network only on synthetic data and demonstrate state-of-the-art performances on real image benchmarks including an extended version of KITTI, PASCAL VOC, PASCAL3D+ and IKEA for 2D and 3D keypoint localization and instance segmentation. The empirical results substantiate the utility of our deep supervision scheme by demonstrating effective transfer of knowledge from synthetic data to real images, resulting in less overfitting compared to standard end-to-end training.

System designers typically use well-studied benchmarks to evaluate and improve new architectures and compilers. We design tomorrow's systems based on yesterday's applications. In this paper we investigate an emerging application, 3D scene understanding, likely to be significant in the mobile space in the near future. Until now, this application could only run in real-time on desktop GPUs. In this work, we examine how it can be mapped to power constrained embedded systems. Key to our approach is the idea of incremental co-design exploration, where optimization choices that concern the domain layer are incrementally explored together with low-level compiler and architecture choices. The goal of this exploration is to reduce execution time while minimizing power and meeting our quality of result objective. As the design space is too large to exhaustively evaluate, we use active learning based on a random forest predictor to find good designs. We show that our approach can, for the first time, achieve dense 3D mapping and tracking in the real-time range within a 1W power budget on a popular embedded device. This is a 4.8x execution time improvement and a 2.8x power reduction compared to the state-of-the-art.

Reasoning about objects in images and videos using 3D representations is re-emerging as a popular paradigm in computer vision. Specifically, in the context of scene understanding for roads, 3D vehicle detection and tracking from monocular videos still needs a lot of attention to enable practical applications. Current approaches leverage two kinds of information to deal with the vehicle detection and tracking problem: (1) 3D representations (eg. wireframe models or voxel based or CAD models) for diverse vehicle skeletal structures learnt from data, and (2) classifiers trained to detect vehicles or vehicle parts in single images built on top of a basic feature extraction step. In this paper, we propose to extend current approaches in two ways. First, we extend detection to a multiple view setting. We show that leveraging information given by feature or part detectors in multiple images can lead to more accurate detection results than single image detection. Secondly, we show that given multiple images of a vehicle, we can also leverage 3D information from the scene generated using a unique structure from motion algorithm. This helps us localize the vehicle in 3D, and constrain the parameters of optimization for fitting the 3D model to image data. We show results on the KITTI dataset, and demonstrate superior results compared with recent state-of-the-art methods, with upto 14.64 % improvement in localization error.

SLAM has matured significantly over the past few years, and is beginning to appear in serious commercial products. While new SLAM systems are being proposed at every conference, evaluation is often restricted to qualitative visualizations or accuracy estimation against a ground truth. This is due to the lack of benchmarking methodologies which can holistically and quantitatively evaluate these systems. Further investigation at the level of individual kernels and parameter spaces of SLAM pipelines is non-existent, which is absolutely essential for systems research and integration. We extend the recently introduced SLAMBench framework to allow comparing two state-of-the-art SLAM pipelines, namely KinectFusion and LSD-SLAM, along the metrics of accuracy, energy consumption, and processing frame rate on two different hardware platforms, namely a desktop and an embedded device. We also analyze the pipelines at the level of individual kernels and explore their algorithmic and hardware design spaces for the first time, yielding valuable insights.

Real-time dense computer vision and SLAM offer great potential for a new level of scene modelling, tracking and real environmental interaction for many types of robot, but their high computational requirements mean that use on mass market embedded platforms is challenging. Meanwhile, trends in low-cost, low-power processing are towards massive parallelism and heterogeneity, making it difficult for robotics and vision researchers to implement their algorithms in a performance-portable way. In this paper we introduce SLAMBench, a publicly-available software framework which represents a starting point for quantitative, comparable and validatable experimental research to investigate trade-offs in performance, accuracy and energy consumption of a dense RGB-D SLAM system. SLAMBench provides a KinectFusion implementation in C++, OpenMP, OpenCL and CUDA, and harnesses the ICL-NUIM dataset of synthetic RGB-D sequences with trajectory and scene ground truth for reliable accuracy comparison of different implementation and algorithms. We present an analysis and breakdown of the constituent algorithmic elements of KinectFusion, and experimentally investigate their execution time on a variety of multicore and GPU-accelerated platforms. For a popular embedded platform, we also present an analysis of energy efficiency for different configuration alternatives.

Current systems for scene understanding typically represent objects as 2D or 3D bounding boxes. While these representations have proven robust in a variety of applications, they provide only coarse approximations to the true 2D and 3D extent of objects. As a result, object-object interactions, such as occlusions or ground-plane contact, can be represented only superficially. In this paper, we approach the problem of scene understanding from the perspective of 3D shape modeling, and design a 3D scene representation that reasons jointly about the 3D shape of multiple objects. This representation allows to express 3D geometry and occlusion on the fine detail level of individual vertices of 3D wireframe models, and makes it possible to treat dependencies between objects, such as occlusion reasoning, in a deterministic way. In our experiments, we demonstrate the benefit of jointly estimating the 3D shape of multiple objects in a scene over working with coarse boxes, on the recently proposed KITTI dataset of realistic street scenes.

Despite the success of current state-of-the-art object class detectors, severe occlusion remains a major challenge. This is particularly true for more geometrically expressive 3D object class representations. While these representations have attracted renewed interest for precise object pose estimation, the focus has mostly been on rather clean datasets, where occlusion is not an issue. In this paper, we tackle the challenge of modeling occlusion in the context of a 3D geometric object class model that is capable of fine-grained, part-level 3D object reconstruction. Following the intuition that 3D modeling should facilitate occlusion reasoning, we design an explicit representation of likely geometric occlusion patterns. Robustness is achieved by pooling image evidence from of a set of fixed part detectors as well as a non-parametric representation of part configurations in the spirit of pose lets. We confirm the potential of our method on cars in a newly collected data set of inner-city street scenes with varying levels of occlusion, and demonstrate superior performance in occlusion estimation and part localization, compared to baselines that are unaware of occlusions.

Geometric 3D reasoning at the level of objects has received renewed attention recently in the context of visual scene understanding. The level of geometric detail, however, is typically limited to qualitative representations or coarse boxes. This is linked to the fact that today's object class detectors are tuned toward robust 2D matching rather than accurate 3D geometry, encouraged by bounding-box-based benchmarks such as Pascal VOC. In this paper, we revisit ideas from the early days of computer vision, namely, detailed, 3D geometric object class representations for recognition. These representations can recover geometrically far more accurate object hypotheses than just bounding boxes, including continuous estimates of object pose and 3D wireframes with relative 3D positions of object parts. In combination with robust techniques for shape description and inference, we outperform state-of-the-art results in monocular 3D pose estimation. In a series of experiments, we analyze our approach in detail and demonstrate novel applications enabled by such an object class representation, such as fine-grained categorization of cars and bicycles, according to their 3D geometry, and ultrawide baseline matching.

We propose a new method for automatically accessing an internet database of 3D models that are searchable only by their user-annotated labels, for using them for vision and robotic manipulation purposes. Instead of having only a local database containing already seen objects, we want to use shared databases available over the internet. This approach while having the potential to dramatically increase the visual recognition capability of robots, also poses certain problems, like wrong annotation due to the open nature of the database, or overwhelming amounts of data (many 3D models) or the lack of relevant data (no models matching a specified label). To solve those problems we propose the following: First, we present an outlier/inlier classification method for reducing the number of results and discarding invalid 3D models that do not match our query. Second, we utilize an approach from computer graphics, the so called ‘morphing’, to this application to specialize the models, in order to describe more objects. Third, we search for 3D models using a restricted search space, as obtained from our knowledge of the environment. We show our classification and matching results and finally show how we can recover the correct scaling with the stereo setup of our robot.