Accurate detection and tracking of objects is vital for effective video understanding. In previous work, the two tasks have been combined in a way that tracking is based heavily on detection, but the detection benefits marginally from the tracking. To increase synergy, we propose to more tightly integrate the tasks by conditioning the object detection in the current frame on tracklets computed in prior frames. With this approach, the object detection results not only have high detection responses, but also improved coherence with the existing tracklets. This greater coherence leads to estimated object trajectories that are smoother and more stable than the jittered paths obtained without tracklet-conditioned detection. Over extensive experiments, this approach is shown to achieve state-of-the-art performance in terms of both detection and tracking accuracy, as well as noticeable improvements in tracking stability.
One of the greatest challenges in the design of a real-time perception system for autonomous driving vehicles and drones is the conflicting requirement of safety (high prediction accuracy) and efficiency. Traditional approaches use a single frame rate for the entire system. Motivated by the observation that the lack of robustness against environmental factors is the major weakness of compact ConvNet architectures, we propose a dual frame-rate system that brings in the best of both worlds: A modulator stream that executes an expensive models robust to environmental factors at a low frame rate to extract slowly changing features describing the environment, and a prediction stream that executes a light-weight model at real-time to extract transient signals that describes particularities of the current frame. The advantage of our design is validated by our extensive empirical study, showing that our solution leads to consistent improvements using a variety of backbone architecture choice and input resolutions. These findings suggest multiple frame-rate systems as a promising direction in designing efficient perception for autonomous agents.
The explosive growth in video streaming gives rise to challenges on performing video understanding at high accuracy and low computation cost. Conventional 2D CNNs are computationally cheap but cannot capture temporal relationships; 3D CNN based methods can achieve good performance but are computationally intensive, making it expensive to deploy. In this paper, we propose a generic and effective Temporal Shift Module (TSM) that enjoys both high efficiency and high performance. Specifically, it can achieve the performance of 3D CNN but maintain 2D CNN's complexity. TSM shifts part of the channels along the temporal dimension; thus facilitate information exchanged among neighboring frames. It can be inserted into 2D CNNs to achieve temporal modeling at zero computation and zero parameters. We also extended TSM to online setting, which enables real-time low-latency online video recognition and video object detection. TSM is accurate and efficient: it ranks the first place on the Something-Something leaderboard upon publication; on Jetson Nano and Galaxy Note8, it achieves a low latency of 13ms and 35ms for online video recognition. The code is available at: https://github.com/mit-han-lab/temporal-shift-module.
This paper introduces a fast and efficient network architecture, NeXtVLAD, to aggregate frame-level features into a compact feature vector for large-scale video classification. Briefly speaking, the basic idea is to decompose a high-dimensional feature into a group of relatively low-dimensional vectors with attention before applying NetVLAD aggregation over time. This NeXtVLAD approach turns out to be both effective and parameter efficient in aggregating temporal information. In the 2nd Youtube-8M video understanding challenge, a single NeXtVLAD model with less than 80M parameters achieves a GAP score of 0.87846 in private leaderboard. A mixture of 3 NeXtVLAD models results in 0.88722, which is ranked 3rd over 394 teams. The code is publicly available at https://github.com/linrongc/youtube-8m.
Current state-of-the-art approaches to video understanding adopt temporal jittering to simulate analyzing the video at varying frame rates. However, this does not work well for multirate videos, in which actions or subactions occur at different speeds. The frame sampling rate should vary in accordance with the different motion speeds. In this work, we propose a simple yet effective strategy, termed random temporal skipping, to address this situation. This strategy effectively handles multirate videos by randomizing the sampling rate during training. It is an exhaustive approach, which can potentially cover all motion speed variations. Furthermore, due to the large temporal skipping, our network can see video clips that originally cover over 100 frames. Such a time range is enough to analyze most actions/events. We also introduce an occlusion-aware optical flow learning method that generates improved motion maps for human action recognition. Our framework is end-to-end trainable, runs in real-time, and achieves state-of-the-art performance on six widely adopted video benchmarks.
The past several years have seen both an explosion in the use of Convolutional Neural Networks (CNNs) and the design of accelerators to make CNN inference practical. In the architecture community, the lion share of effort has targeted CNN inference for image recognition. The closely related problem of video recognition has received far less attention as an accelerator target. This is surprising, as video recognition is more computationally intensive than image recognition, and video traffic is predicted to be the majority of internet traffic in the coming years. This paper fills the gap between algorithmic and hardware advances for video recognition by providing a design space exploration and flexible architecture for accelerating 3D Convolutional Neural Networks (3D CNNs) - the core kernel in modern video understanding. When compared to (2D) CNNs used for image recognition, efficiently accelerating 3D CNNs poses a significant engineering challenge due to their large (and variable over time) memory footprint and higher dimensionality. To address these challenges, we design a novel accelerator, called Morph, that can adaptively support different spatial and temporal tiling strategies depending on the needs of each layer of each target 3D CNN. We codesign a software infrastructure alongside the Morph hardware to find good-fit parameters to control the hardware. Evaluated on state-of-the-art 3D CNNs, Morph achieves up to 3.4x (2.5x average) reduction in energy consumption and improves performance/watt by up to 5.1x (4x average) compared to a baseline 3D CNN accelerator, with an area overhead of 5%. Morph further achieves a 15.9x average energy reduction on 3D CNNs when compared to Eyeriss.
In this paper, we propose a convolutional layer inspired by optical flow algorithms to learn motion representations. Our representation flow layer is a fully-differentiable layer designed to capture the `flow' of any representation channel within a convolutional neural network for action recognition. Its parameters for iterative flow optimization are learned in an end-to-end fashion together with the other CNN model parameters, maximizing the action recognition performance. Furthermore, we newly introduce the concept of learning `flow of flow' representations by stacking multiple representation flow layers. We conducted extensive experimental evaluations, confirming its advantages over previous recognition models using traditional optical flows in both computational speed and performance. Code/models available here: https://piergiaj.github.io/rep-flow-site/
Deep learning based object detectors require thousands of diversified bounding box and class annotated examples. Though image object detectors have shown rapid progress in recent years with the release of multiple large-scale static image datasets, object detection on videos still remains an open problem due to scarcity of annotated video frames. Having a robust video object detector is an essential component for video understanding and curating large-scale automated annotations in videos. Domain difference between images and videos makes the transferability of image object detectors to videos sub-optimal. The most common solution is to use weakly supervised annotations where a video frame has to be tagged for presence/absence of object categories. This still takes up manual effort. In this paper we take a step forward by adapting the concept of unsupervised adversarial image-to-image translation to perturb static high quality images to be visually indistinguishable from a set of video frames. We assume the presence of a fully annotated static image dataset and an unannotated video dataset. Object detector is trained on adversarially transformed image dataset using the annotations of the original dataset. Experiments on Youtube-Objects and Youtube-Objects-Subset datasets with two contemporary baseline object detectors reveal that such unsupervised pixel level domain adaptation boosts the generalization performance on video frames compared to direct application of original image object detector. Also, we achieve competitive performance compared to recent baselines of weakly supervised methods. This paper can be seen as an application of image translation for cross domain object detection.
This paper describes our solution for the 2$^\text{nd}$ YouTube-8M video understanding challenge organized by Google AI. Unlike the video recognition benchmarks, such as Kinetics and Moments, the YouTube-8M challenge provides pre-extracted visual and audio features instead of raw videos. In this challenge, the submitted model is restricted to 1GB, which encourages participants focus on constructing one powerful single model rather than incorporating of the results from a bunch of models. Our system fuses six different sub-models into one single computational graph, which are categorized into three families. More specifically, the most effective family is the model with non-local operations following the NetVLAD encoding. The other two family models are Soft-BoF and GRU, respectively. In order to further boost single models performance, the model parameters of different checkpoints are averaged. Experimental results demonstrate that our proposed system can effectively perform the video classification task, achieving 0.88763 on the public test set and 0.88704 on the private set in terms of GAP@20, respectively. We finally ranked at the fourth place in the YouTube-8M video understanding challenge.
We introduce modifications to state-of-the-art approaches to aggregating local video descriptors by using attention mechanisms and function approximations. Rather than using ensembles of existing architectures, we provide an insight on creating new architectures. We demonstrate our solutions in the "The 2nd YouTube-8M Video Understanding Challenge", by using frame-level video and audio descriptors. We obtain testing accuracy similar to the state of the art, while meeting budget constraints, and touch upon strategies to improve the state of the art. Model implementations are available in https://github.com/pomonam/LearnablePoolingMethods.
This paper presents the Axon AI's solution to the 2nd YouTube-8M Video Understanding Challenge, achieving the final global average precision (GAP) of 88.733% on the private test set (ranked 3rd among 394 teams, not considering the model size constraint), and 87.287% using a model that meets size requirement. Two sets of 7 individual models belonging to 3 different families were trained separately. Then, the inference results on a training data were aggregated from these multiple models and fed to train a compact model that meets the model size requirement. In order to further improve performance we explored and employed data over/sub-sampling in feature space, an additional regularization term during training exploiting label relationship, and learned weights for ensembling different individual models.
Despite recent advances in computer vision based on various convolutional architectures, video understanding remains an important challenge. In this work, we present and discuss a top solution for the large-scale video classification (labeling) problem introduced as a Kaggle competition based on the YouTube-8M dataset. We show and compare different approaches to preprocessing, data augmentation, model architectures, and model combination. Our final model is based on a large ensemble of video- and frame-level models but fits into rather limiting hardware constraints. We apply an approach based on knowledge distillation to deal with noisy labels in the original dataset and the recently developed mixup technique to improve the basic models.
Localizing moments in a longer video via natural language queries is a new, challenging task at the intersection of language and video understanding. Though moment localization with natural language is similar to other language and vision tasks like natural language object retrieval in images, moment localization offers an interesting opportunity to model temporal dependencies and reasoning in text. We propose a new model that explicitly reasons about different temporal segments in a video, and shows that temporal context is important for localizing phrases which include temporal language. To benchmark whether our model, and other recent video localization models, can effectively reason about temporal language, we collect the novel TEMPOral reasoning in video and language (TEMPO) dataset. Our dataset consists of two parts: a dataset with real videos and template sentences (TEMPO - Template Language) which allows for controlled studies on temporal language, and a human language dataset which consists of temporal sentences annotated by humans (TEMPO - Human Language).
This paper presents our 7th place solution to the second YouTube-8M video understanding competition which challenges participates to build a constrained-size model to classify millions of YouTube videos into thousands of classes. Our final model consists of four single models aggregated into one tensorflow graph. For each single model, we use the same network architecture as in the winning solution of the first YouTube-8M video understanding competition, namely Gated NetVLAD. We train the single models separately in tensorflow's default float32 precision, then replace weights with float16 precision and ensemble them in the evaluation and inference stages., achieving 48.5% compression rate without loss of precision. Our best model achieved 88.324% GAP on private leaderboard. The code is publicly available at https://github.com/boliu61/youtube-8m
Despite the recent progress in video understanding and the continuous rate of improvement in temporal action localization throughout the years, it is still unclear how far (or close?) we are to solving the problem. To this end, we introduce a new diagnostic tool to analyze the performance of temporal action detectors in videos and compare different methods beyond a single scalar metric. We exemplify the use of our tool by analyzing the performance of the top rewarded entries in the latest ActivityNet action localization challenge. Our analysis shows that the most impactful areas to work on are: strategies to better handle temporal context around the instances, improving the robustness w.r.t. the instance absolute and relative size, and strategies to reduce the localization errors. Moreover, our experimental analysis finds the lack of agreement among annotator is not a major roadblock to attain progress in the field. Our diagnostic tool is publicly available to keep fueling the minds of other researchers with additional insights about their algorithms.
Video summarization plays an important role in video understanding by selecting key frames/shots. Traditionally, it aims to find the most representative and diverse contents in a video as short summaries. Recently, a more generalized task, query-conditioned video summarization, has been introduced, which takes user queries into consideration to learn more user-oriented summaries. In this paper, we propose a query-conditioned three-player generative adversarial network to tackle this challenge. The generator learns the joint representation of the user query and the video content, and the discriminator takes three pairs of query-conditioned summaries as the input to discriminate the real summary from a generated and a random one. A three-player loss is introduced for joint training of the generator and the discriminator, which forces the generator to learn better summary results, and avoids the generation of random trivial summaries. Experiments on a recently proposed query-conditioned video summarization benchmark dataset show the efficiency and efficacy of our proposed method.
Time-aware encoding of frame sequences in a video is a fundamental problem in video understanding. While many attempted to model time in videos, an explicit study on quantifying video time is missing. To fill this lacuna, we aim to evaluate video time explicitly. We describe three properties of video time, namely a) temporal asymmetry, b)temporal continuity and c) temporal causality. Based on each we formulate a task able to quantify the associated property. This allows assessing the effectiveness of modern video encoders, like C3D and LSTM, in their ability to model time. Our analysis provides insights about existing encoders while also leading us to propose a new video time encoder, which is better suited for the video time recognition tasks than C3D and LSTM. We believe the proposed meta-analysis can provide a reasonable baseline to assess video time encoders on equal grounds on a set of temporal-aware tasks.
In this work we introduce a time- and memory-efficient method for structured prediction that couples neuron decisions across both space at time. We show that we are able to perform exact and efficient inference on a densely connected spatio-temporal graph by capitalizing on recent advances on deep Gaussian Conditional Random Fields (GCRFs). Our method, called VideoGCRF is (a) efficient, (b) has a unique global minimum, and (c) can be trained end-to-end alongside contemporary deep networks for video understanding. We experiment with multiple connectivity patterns in the temporal domain, and present empirical improvements over strong baselines on the tasks of both semantic and instance segmentation of videos.
Numerous pattern recognition applications can be formed as learning from graph-structured data, including social network, protein-interaction network, the world wide web data, knowledge graph, etc. While convolutional neural network (CNN) facilitates great advances in gridded image/video understanding tasks, very limited attention has been devoted to transform these successful network structures (including Inception net, Residual net, Dense net, etc.) to establish convolutional networks on graph, due to its irregularity and complexity geometric topologies (unordered vertices, unfixed number of adjacent edges/vertices). In this paper, we aim to give a comprehensive analysis of when work matters by transforming different classical network structures to graph CNN, particularly in the basic graph recognition problem. Specifically, we firstly review the general graph CNN methods, especially in its spectral filtering operation on the irregular graph data. We then introduce the basic structures of ResNet, Inception and DenseNet into graph CNN and construct these network structures on graph, named as G_ResNet, G_Inception, G_DenseNet. In particular, it seeks to help graph CNNs by shedding light on how these classical network structures work and providing guidelines for choosing appropriate graph network frameworks. Finally, we comprehensively evaluate the performance of these different network structures on several public graph datasets (including social networks and bioinformatic datasets), and demonstrate how different network structures work on graph CNN in the graph recognition task.
Video understanding is one of the most challenging topics in computer vision. In this paper, a four-stage video understanding pipeline is presented to simultaneously recognize all atomic actions and the single on-going activity in a video. This pipeline uses objects and motions from the video and a graph-based knowledge representation network as prior reference. Two deep networks are trained to identify objects and motions in each video sequence associated with an action. Low Level image features are then used to identify objects of interest in that video sequence. Confidence scores are assigned to objects of interest based on their involvement in the action and to motion classes based on results from a deep neural network that classifies the on-going action in video into motion classes. Confidence scores are computed for each candidate functional unit associated with an action using a knowledge representation network, object confidences, and motion confidences. Each action is therefore associated with a functional unit and the sequence of actions is further evaluated to identify the single on-going activity in the video. The knowledge representation used in the pipeline is called the functional object-oriented network which is a graph-based network useful for encoding knowledge about manipulation tasks. Experiments are performed on a dataset of cooking videos to test the proposed algorithm with action inference and activity classification. Experiments show that using functional object oriented network improves video understanding significantly.