In this thesis, we focus on video action understanding problems from an online and real-time processing point of view. We start with the conversion of the traditional offline spatiotemporal action detection pipeline into an online spatiotemporal action tube detection system. An action tube is a set of bounding connected over time, which bounds an action instance in space and time. Next, we explore the future prediction capabilities of such detection methods by extending an existing action tube into the future by regression. Later, we seek to establish that online/causal representations can achieve similar performance to that of offline three dimensional (3D) convolutional neural networks (CNNs) on various tasks, including action recognition, temporal action segmentation and early prediction.
We propose an effective framework for the temporal action segmentation task, namely an Action Segment Refinement Framework (ASRF). Our model architecture consists of a long-term feature extractor and two branches: the Action Segmentation Branch (ASB) and the Boundary Regression Branch (BRB). The long-term feature extractor provides shared features for the two branches with a wide temporal receptive field. The ASB classifies video frames with action classes, while the BRB regresses the action boundary probabilities. The action boundaries predicted by the BRB refine the output from the ASB, which results in a significant performance improvement. Our contributions are three-fold: (i) We propose a framework for temporal action segmentation, the ASRF, which divides temporal action segmentation into frame-wise action classification and action boundary regression. Our framework refines frame-level hypotheses of action classes using predicted action boundaries. (ii) We propose a loss function for smoothing the transition of action probabilities, and analyze combinations of various loss functions for temporal action segmentation. (iii) Our framework outperforms state-of-the-art methods on three challenging datasets, offering an improvement of up to 13.7% in terms of segmental edit distance and up to 16.1% in terms of segmental F1 score. Our code will be publicly available soon.
With the success of deep learning in classifying short trimmed videos, more attention has been focused on temporally segmenting and classifying activities in long untrimmed videos. State-of-the-art approaches for action segmentation utilize several layers of temporal convolution and temporal pooling. Despite the capabilities of these approaches in capturing temporal dependencies, their predictions suffer from over-segmentation errors. In this paper, we propose a multi-stage architecture for the temporal action segmentation task that overcomes the limitations of the previous approaches. The first stage generates an initial prediction that is refined by the next ones. In each stage we stack several layers of dilated temporal convolutions covering a large receptive field with few parameters. While this architecture already performs well, lower layers still suffer from a small receptive field. To address this limitation, we propose a dual dilated layer that combines both large and small receptive fields. We further decouple the design of the first stage from the refining stages to address the different requirements of these stages. Extensive evaluation shows the effectiveness of the proposed model in capturing long-range dependencies and recognizing action segments. Our models achieve state-of-the-art results on three datasets: 50Salads, Georgia Tech Egocentric Activities (GTEA), and the Breakfast dataset.
Learning meaningful visual representations in an embedding space can facilitate generalization in downstream tasks such as action segmentation and imitation. In this paper, we learn a motion-centric representation of surgical video demonstrations by grouping them into action segments/sub-goals/options in a semi-supervised manner. We present Motion2Vec, an algorithm that learns a deep embedding feature space from video observations by minimizing a metric learning loss in a Siamese network: images from the same action segment are pulled together while pushed away from randomly sampled images of other segments, while respecting the temporal ordering of the images. The embeddings are iteratively segmented with a recurrent neural network for a given parametrization of the embedding space after pre-training the Siamese network. We only use a small set of labeled video segments to semantically align the embedding space and assign pseudo-labels to the remaining unlabeled data by inference on the learned model parameters. We demonstrate the use of this representation to imitate surgical suturing motions from publicly available videos of the JIGSAWS dataset. Results give 85.5 % segmentation accuracy on average suggesting performance improvement over several state-of-the-art baselines, while kinematic pose imitation gives 0.94 centimeter error in position per observation on the test set. Videos, code and data are available at https://sites.google.com/view/motion2vec
Action segmentation is the task of temporally segmenting every frame of an untrimmed video. Weakly supervised approaches to action segmentation, especially from transcripts have been of considerable interest to the computer vision community. In this work, we focus on two aspects of the use and evaluation of weakly supervised action segmentation approaches that are often overlooked: the performance variance over multiple training runs and the impact of selecting feature extractors for this task. To tackle the first problem, we train each method on the Breakfast dataset 5 times and provide average and standard deviation of the results. Our experiments show that the standard deviation over these repetitions is between 1 and 2.5% and significantly affects the comparison between different approaches. Furthermore, our investigation on feature extraction shows that, for the studied weakly-supervised action segmentation methods, higher-level I3D features perform worse than classical IDT features.
We apply a generative segmental model of task structure, guided by narration, to action segmentation in video. We focus on unsupervised and weakly-supervised settings where no action labels are known during training. Despite its simplicity, our model performs competitively with previous work on a dataset of naturalistic instructional videos. Our model allows us to vary the sources of supervision used in training, and we find that both task structure and narrative language provide large benefits in segmentation quality.
The temporal segmentation of events is an essential task and a precursor for the automatic recognition of human actions in the video. Several attempts have been made to capture frame-level salient aspects through attention but they lack the capacity to effectively map the temporal relationships in between the frames as they only capture a limited span of temporal dependencies. To this end we propose a complete end-to-end supervised learning approach that can better learn relationships between actions over time, thus improving the overall segmentation performance. The proposed hierarchical recurrent attention framework analyses the input video at multiple temporal scales, to form embeddings at frame level and segment level, and perform fine-grained action segmentation. This generates a simple, lightweight, yet extremely effective architecture for segmenting continuous video streams and has multiple application domains. We evaluate our system on multiple challenging public benchmark datasets, including MERL Shopping, 50 salads, and Georgia Tech Egocentric datasets, and achieves state-of-the-art performance. The evaluated datasets encompass numerous video capture settings which are inclusive of static overhead camera views and dynamic, ego-centric head-mounted camera views, demonstrating the direct applicability of the proposed framework in a variety of settings.
Temporal action segmentation is a topic of increasing interest, however, annotating each frame in a video is cumbersome and costly. Weakly supervised approaches therefore aim at learning temporal action segmentation from videos that are only weakly labeled. In this work, we assume that for each training video only the list of actions is given that occur in the video, but not when, how often, and in which order they occur. In order to address this task, we propose an approach that can be trained end-to-end on such data. The approach divides the video into smaller temporal regions and predicts for each region the action label and its length. In addition, the network estimates the action labels for each frame. By measuring how consistent the frame-wise predictions are with respect to the temporal regions and the annotated action labels, the network learns to divide a video into class-consistent regions. We evaluate our approach on three datasets where the approach achieves state-of-the-art results.
We address weakly-supervised video actor-action segmentation (VAAS), which extends general video object segmentation (VOS) to additionally consider action labels of the actors. The most successful methods on VOS synthesize a pool of pseudo-annotations (PAs) and then refine them iteratively. However, they face challenges as to how to select from a massive amount of PAs high-quality ones, how to set an appropriate stop condition for weakly-supervised training, and how to initialize PAs pertaining to VAAS. To overcome these challenges, we propose a general Weakly-Supervised framework with a Wise Selection of training samples and model evaluation criterion (WS^2). Instead of blindly trusting quality-inconsistent PAs, WS^2 employs a learning-based selection to select effective PAs and a novel region integrity criterion as a stopping condition for weakly-supervised training. In addition, a 3D-Conv GCAM is devised to adapt to the VAAS task. Extensive experiments show that WS^2 achieves state-of-the-art performance on both weakly-supervised VOS and VAAS tasks and is on par with the best fully-supervised method on VAAS.
Despite the recent progress of fully-supervised action segmentation techniques, the performance is still not fully satisfactory. One main challenge is the problem of spatiotemporal variations (e.g. different people may perform the same activity in various ways). Therefore, we exploit unlabeled videos to address this problem by reformulating the action segmentation task as a cross-domain problem with domain discrepancy caused by spatio-temporal variations. To reduce the discrepancy, we propose Self-Supervised Temporal Domain Adaptation (SSTDA), which contains two self-supervised auxiliary tasks (binary and sequential domain prediction) to jointly align cross-domain feature spaces embedded with local and global temporal dynamics, achieving better performance than other Domain Adaptation (DA) approaches. On three challenging benchmark datasets (GTEA, 50Salads, and Breakfast), SSTDA outperforms the current state-of-the-art method by large margins (e.g. for the F1@25 score, from 59.6% to 69.1% on Breakfast, from 73.4% to 81.5% on 50Salads, and from 83.6% to 89.1% on GTEA), and requires only 65% of the labeled training data for comparable performance, demonstrating the usefulness of adapting to unlabeled target videos across variations. The source code is available at https://github.com/cmhungsteve/SSTDA.
This paper is about weakly supervised action segmentation, where the ground truth specifies only a set of actions present in a training video, but not their true temporal ordering. Prior work typically uses a classifier that independently labels video frames for generating the pseudo ground truth, and multiple instance learning for training the classifier. We extend this framework by specifying an HMM, which accounts for co-occurrences of action classes and their temporal lengths, and by explicitly training the HMM on a Viterbi-based loss. Our first contribution is the formulation of a new set-constrained Viterbi algorithm (SCV). Given a video, the SCV generates the MAP action segmentation that satisfies the ground truth. This prediction is used as a framewise pseudo ground truth in our HMM training. Our second contribution in training is a new regularization of feature affinities between training videos that share the same action classes. Evaluation on action segmentation and alignment on the Breakfast, MPII Cooking2, Hollywood Extended datasets demonstrates our significant performance improvement for the two tasks over prior work.
Automatic surgical gesture recognition is fundamental for improving intelligence in robot-assisted surgery, such as conducting complicated tasks of surgery surveillance and skill evaluation. However, current methods treat each frame individually and produce the outcomes without effective consideration on future information. In this paper, we propose a framework based on reinforcement learning and tree search for joint surgical gesture segmentation and classification. An agent is trained to segment and classify the surgical video in a human-like manner whose direct decisions are re-considered by tree search appropriately. Our proposed tree search algorithm unites the outputs from two designed neural networks, i.e., policy and value network. With the integration of complementary information from distinct models, our framework is able to achieve the better performance than baseline methods using either of the neural networks. For an overall evaluation, our developed approach consistently outperforms the existing methods on the suturing task of JIGSAWS dataset in terms of accuracy, edit score and F1 score. Our study highlights the utilization of tree search to refine actions in reinforcement learning framework for surgical robotic applications.
With the recent success of the pre-training technique for NLP and image-linguistic tasks, some video-linguistic pre-training works are gradually developed to improve video-text related downstream tasks. However, most of the existing multimodal models are pre-trained for understanding tasks, leading to a pretrain-finetune discrepancy for generation tasks. This paper proposes UniVL: a Unified Video and Language pre-training model for both multimodal understanding and generation. It comprises four components, including two single-modal encoders, a cross encoder, and a decoder with the Transformer backbone. Five objectives, including video-text joint, conditioned masked language model (CMLM), conditioned masked frame model (CMFM), video-text alignment, and language reconstruction, are designed to train each of the components. We further develop two pre-training strategies, stage by stage pre-training (StagedP) and enhanced video representation (EnhancedV), to make the training process of the UniVL more effective. The pre-train is carried out on a sizeable instructional video dataset HowTo100M. Experimental results demonstrate that the UniVL can learn strong video-text representation and achieves state-of-the-art results on five downstream tasks.
Understanding the structure of complex activities in untrimmed videos is a challenging task in the area of action recognition. One problem here is that this task usually requires a large amount of hand-annotated minute- or even hour-long video data, but annotating such data is very time consuming and can not easily be automated or scaled. To address this problem, this paper proposes an approach for the unsupervised learning of actions in untrimmed video sequences based on a joint visual-temporal embedding space. To this end, we combine a visual embedding based on a predictive U-Net architecture with a temporal continuous function. The resulting representation space allows detecting relevant action clusters based on their visual as well as their temporal appearance. The proposed method is evaluated on three standard benchmark datasets, Breakfast Actions, INRIA YouTube Instructional Videos, and 50 Salads. We show that the proposed approach is able to provide a meaningful visual and temporal embedding out of the visual cues present in contiguous video frames and is suitable for the task of unsupervised temporal segmentation of actions.
Annotating videos is cumbersome, expensive and not scalable. Yet, many strong video models still rely on manually annotated data. With the recent introduction of the HowTo100M dataset, narrated videos now offer the possibility of learning video representations without manual supervision. In this work we propose a new learning approach, MIL-NCE, capable of addressing misalignments inherent to narrated videos. With this approach we are able to learn strong video representations from scratch, without the need for any manual annotation. We evaluate our representations on a wide range of four downstream tasks over eight datasets: action recognition (HMDB-51, UCF-101, Kinetics-700), text-to-video retrieval (YouCook2, MSR-VTT), action localization (YouTube-8M Segments, CrossTask) and action segmentation (COIN). Our method outperforms all published self-supervised approaches for these tasks as well as several fully supervised baselines.
This paper classifies human action sequences from videos using a machine translation model. In contrast to classical human action classification which outputs a set of actions, our method output a sequence of action in the chronological order of the actions performed by the human. Therefore our method is evaluated using sequential performance measures such as Bilingual Evaluation Understudy (BLEU) scores. Action sequence classification has many applications such as learning from demonstration, action segmentation, detection, localization and video captioning. Furthermore, we use our model that is trained to output action sequences to solve downstream tasks; such as video captioning and action localization. We obtain state of the art results for video captioning in challenging Charades dataset obtaining BLEU-4 score of 34.8 and METEOR score of 33.6 outperforming previous state-of-the-art of 18.8 and 19.5 respectively. Similarly, on ActivityNet captioning, we obtain excellent results in-terms of ROUGE (20.24) and CIDER (37.58) scores. For action localization, without using any explicit start/end action annotations, our method obtains localization performance of 22.2 mAP outperforming prior fully supervised methods.
This paper is about labeling video frames with action classes under weak supervision in training, where we have access to a temporal ordering of actions, but their start and end frames in training videos are unknown. Following prior work, we use an HMM grounded on a Gated Recurrent Unit (GRU) for frame labeling. Our key contribution is a new constrained discriminative forward loss (CDFL) that we use for training the HMM and GRU under weak supervision. While prior work typically estimates the loss on a single, inferred video segmentation, our CDFL discriminates between the energy of all valid and invalid frame labelings of a training video. A valid frame labeling satisfies the ground-truth temporal ordering of actions, whereas an invalid one violates the ground truth. We specify an efficient recursive algorithm for computing the CDFL in terms of the logadd function of the segmentation energy. Our evaluation on action segmentation and alignment gives superior results to those of the state of the art on the benchmark Breakfast Action, Hollywood Extended, and 50Salads datasets.
In this paper we address the problem of continuous fine-grained action segmentation, in which multiple actions are present in an unsegmented video stream. The challenge for this task lies in the need to represent the hierarchical nature of the actions and to detect the transitions between actions, allowing us to localise the actions within the video effectively. We propose a novel recurrent semi-supervised Generative Adversarial Network (GAN) model for continuous fine-grained human action segmentation. Temporal context information is captured via a novel Gated Context Extractor (GCE) module, composed of gated attention units, that directs the queued context information through the generator model, for enhanced action segmentation. The GAN is made to learn features in a semi-supervised manner, enabling the model to perform action classification jointly with the standard, unsupervised, GAN learning procedure. We perform extensive evaluations on different architectural variants to demonstrate the importance of the proposed network architecture, and show that it is capable of outperforming current state-of-the-art on three challenging datasets: 50 Salads, MERL Shopping and Georgia Tech Egocentric Activities dataset.
We propose a novel conditional GAN (cGAN) model for continuous fine-grained human action segmentation, that utilises multi-modal data and learned scene context information. The proposed approach utilises two GANs: termed Action GAN and Auxiliary GAN, where the Action GAN is trained to operate over the current RGB frame while the Auxiliary GAN utilises supplementary information such as depth or optical flow. The goal of both GANs is to generate similar `action codes', a vector representation of the current action. To facilitate this process a context extractor that incorporates data and recent outputs from both modes is used to extract context information to aid recognition. The result is a recurrent GAN architecture which learns a task specific loss function from multiple feature modalities. Extensive evaluations on variants of the proposed model to show the importance of utilising different information streams such as context and auxiliary information in the proposed network; and show that our model is capable of outperforming state-of-the-art methods for three widely used datasets: 50 Salads, MERL Shopping and Georgia Tech Egocentric Activities, comprising both static and dynamic camera settings.
Convolutional Neural Network (CNN) based image segmentation has made great progress in recent years. However, video object segmentation remains a challenging task due to its high computational complexity. Most of the previous methods employ a two-stream CNN framework to handle spatial and motion features separately. In this paper, we propose an end-to-end encoder-decoder style 3D CNN to aggregate spatial and temporal information simultaneously for video object segmentation. To efficiently process video, we propose 3D separable convolution for the pyramid pooling module and decoder, which dramatically reduces the number of operations while maintaining the performance. Moreover, we also extend our framework to video action segmentation by adding an extra classifier to predict the action label for actors in videos. Extensive experiments on several video datasets demonstrate the superior performance of the proposed approach for action and object segmentation compared to the state-of-the-art.