This technical report describes the CONE approach for Ego4D Natural Language Queries (NLQ) Challenge in ECCV 2022. We leverage our model CONE, an efficient window-centric COarse-to-fiNE alignment framework. Specifically, CONE dynamically slices the long video into candidate windows via a sliding window approach. Centering at windows, CONE (1) learns the inter-window (coarse-grained) semantic variance through contrastive learning and speeds up inference by pre-filtering the candidate windows relevant to the NL query, and (2) conducts intra-window (fine-grained) candidate moments ranking utilizing the powerful multi-modal alignment ability of the contrastive vision-text pre-trained model EgoVLP. On the blind test set, CONE achieves 15.26 and 9.24 for R1@IoU=0.3 and R1@IoU=0.5, respectively.
Video temporal grounding (VTG) targets to localize temporal moments in an untrimmed video according to a natural language (NL) description. Since real-world applications provide a never-ending video stream, it raises demands for temporal grounding for long-form videos, which leads to two major challenges: (1) the long video length makes it difficult to process the entire video without decreasing sample rate and leads to high computational burden; (2) the accurate multi-modal alignment is more challenging as the number of moment candidates increases. To address these challenges, we propose CONE, an efficient window-centric COarse-to-fiNE alignment framework, which flexibly handles long-form video inputs with higher inference speed, and enhances the temporal grounding via our novel coarse-to-fine multi-modal alignment framework. Specifically, we dynamically slice the long video into candidate windows via a sliding window approach. Centering at windows, CONE (1) learns the inter-window (coarse-grained) semantic variance through contrastive learning and speeds up inference by pre-filtering the candidate windows relevant to the NL query, and (2) conducts intra-window (fine-grained) candidate moments ranking utilizing the powerful multi-modal alignment ability of a contrastive vision-text pre-trained model. Extensive experiments on two large-scale VTG benchmarks for long videos consistently show a substantial performance gain (from 3.13% to 6.87% on MAD and from 10.46% to 13.46% on Ego4d-NLQ) and CONE achieves the SOTA results on both datasets. Analysis reveals the effectiveness of components and higher efficiency in long video grounding as our system improves the inference speed by 2x on Ego4d-NLQ and 15x on MAD while keeping the SOTA performance of CONE.
Conventional video models rely on a single stream to capture the complex spatial-temporal features. Recent work on two-stream video models, such as SlowFast network and AssembleNet, prescribe separate streams to learn complementary features, and achieve stronger performance. However, manually designing both streams as well as the in-between fusion blocks is a daunting task, requiring to explore a tremendously large design space. Such manual exploration is time-consuming and often ends up with sub-optimal architectures when computational resources are limited and the exploration is insufficient. In this work, we present a pragmatic neural architecture search approach, which is able to search for two-stream video models in giant spaces efficiently. We design a multivariate search space, including 6 search variables to capture a wide variety of choices in designing two-stream models. Furthermore, we propose a progressive search procedure, by searching for the architecture of individual streams, fusion blocks, and attention blocks one after the other. We demonstrate two-stream models with significantly better performance can be automatically discovered in our design space. Our searched two-stream models, namely Auto-TSNet, consistently outperform other models on standard benchmarks. On Kinetics, compared with the SlowFast model, our Auto-TSNet-L model reduces FLOPS by nearly 11 times while achieving the same accuracy 78.9%. On Something-Something-V2, Auto-TSNet-M improves the accuracy by at least 2% over other methods which use less than 50 GFLOPS per video.
Localizing persons and recognizing their actions from videos is a challenging task towards high-level video understanding. Recent advances have been achieved by modeling either 'actor-actor' or 'actor-context' relations. However, such direct first-order relations are not sufficient for localizing actions in complicated scenes. Some actors might be indirectly related via objects or background context in the scene. Such indirect relations are crucial for determining the action labels but are mostly ignored by existing work. In this paper, we propose to explicitly model the Actor-Context-Actor Relation, which can capture indirect high-order supportive information for effectively reasoning actors' actions in complex scenes. To this end, we design an Actor-Context-Actor Relation Network (ACAR-Net) which builds upon a novel High-order Relation Reasoning Operator to model indirect relations for spatio-temporal action localization. Moreover, to allow utilizing more temporal contexts, we extend our framework with an Actor-Context Feature Bank for reasoning long-range high-order relations. Extensive experiments on AVA dataset validate the effectiveness of our ACAR-Net. Ablation studies show the advantages of modeling high-order relations over existing first-order relation reasoning methods. The proposed ACAR-Net is also the core module of our 1st place solution in AVA-Kinetics Crossover Challenge 2020. Training code and models will be available at https://github.com/Siyu-C/ACAR-Net.
In this paper, we study an intermediate form of supervision, i.e., single-frame supervision, for temporal action localization (TAL). To obtain the single-frame supervision, the annotators are asked to identify only a single frame within the temporal window of an action. This can significantly reduce the labor cost of obtaining full supervision which requires annotating the action boundary. Compared to the weak supervision that only annotates the video-level label, the single-frame supervision introduces extra temporal action signals while maintaining low annotation overhead. To make full use of such single-frame supervision, we propose a unified system called SF-Net. First, we propose to predict an actionness score for each video frame. Along with a typical category score, the actionness score can provide comprehensive information about the occurrence of a potential action and aid the temporal boundary refinement during inference. Second, we mine pseudo action and background frames based on the single-frame annotations. We identify pseudo action frames by adaptively expanding each annotated single frame to its nearby, contextual frames and we mine pseudo background frames from all the unannotated frames across multiple videos. Together with the ground-truth labeled frames, these pseudo-labeled frames are further used for training the classifier. In extensive experiments on THUMOS14, GTEA, and BEOID, SF-Net significantly improves upon state-of-the-art weakly-supervised methods in terms of both segment localization and single-frame localization. Notably, SF-Net achieves comparable results to its fully-supervised counterpart which requires much more resource intensive annotations.
Recently, Weakly-supervised Temporal Action Localization (WTAL) has been densely studied because it can free us from costly annotating temporal boundaries of actions. One prevalent strategy is obtaining action score sequences over time and then truncating segments of scores higher than a fixed threshold at every kept snippet. However, the threshold is not modeled in the training process and manually setting the threshold introduces expert knowledge, which damages the coherence of systems and makes it unfair for comparisons. In this paper, we propose to adaptively set the threshold at each snippet to be its background score, which can be learned to predict (LPAT). In both training and testing time, the predicted threshold is leveraged to localize action segments and the scores of these segments are allocated for video classification. We also identify an important constraint to improve the confidence of generated proposals, and model it as a novel loss term, which facilitates the video classification loss to improve models' localization ability. As such, our LPAT model is able to generate accurate action proposals with only video-level supervision. Extensive experiments on two standard yet challenging datasets, i.e., THUMOS'14 and ActivityNet1.2, show significant improvement over state-of-the-art methods.