Papers

Due to difficulties in acquiring ground truth depth of equirectangular (360) images, the quality and quantity of equirectangular depth data today is insufficient to represent the various scenes in the world. Therefore, 360 depth estimation studies, which relied solely on supervised learning, are destined to produce unsatisfactory results. Although self-supervised learning methods focusing on equirectangular images (EIs) are introduced, they often have incorrect or non-unique solutions, causing unstable performance. In this paper, we propose 360 monocular depth estimation methods which improve on the areas that limited previous studies. First, we introduce a self-supervised 360 depth learning method that only utilizes gravity-aligned videos, which has the potential to eliminate the needs for depth data during the training procedure. Second, we propose a joint learning scheme realized by combining supervised and self-supervised learning. The weakness of each learning is compensated, thus leading to more accurate depth estimation. Third, we propose a non-local fusion block, which can further retain the global information encoded by vision transformer when reconstructing the depths. With the proposed methods, we successfully apply the transformer to 360 depth estimations, to the best of our knowledge, which has not been tried before. On several benchmarks, our approach achieves significant improvements over previous works and establishes a state of the art.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Spatiotemporal modeling in an unified architecture is key for video action recognition. This paper proposes a Shrinking Temporal Attention Transformer (STAT), which efficiently builts spatiotemporal attention maps considering the attenuation of spatial attention in short and long temporal sequences. Specifically, for short-term temporal tokens, query token interacts with them in a fine-grained manner in dealing with short-range motion. It then shrinks to a coarse attention in neighborhood for long-term tokens, to provide larger receptive field for long-range spatial aggregation. Both of them are composed in a short-long temporal integrated block to build visual appearances and temporal structure concurrently with lower costly in computation. We conduct thorough ablation studies, and achieve state-of-the-art results on multiple action recognition benchmarks including Kinetics400 and Something-Something v2, outperforming prior methods with 50% less FLOPs and without any pretrained model.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Despite the great progress in video understanding made by deep convolutional neural networks, feature representation learned by existing methods may be biased to static visual cues. To address this issue, we propose a novel method to suppress static visual cues (SSVC) based on probabilistic analysis for self-supervised video representation learning. In our method, video frames are first encoded to obtain latent variables under standard normal distribution via normalizing flows. By modelling static factors in a video as a random variable, the conditional distribution of each latent variable becomes shifted and scaled normal. Then, the less-varying latent variables along time are selected as static cues and suppressed to generate motion-preserved videos. Finally, positive pairs are constructed by motion-preserved videos for contrastive learning to alleviate the problem of representation bias to static cues. The less-biased video representation can be better generalized to various downstream tasks. Extensive experiments on publicly available benchmarks demonstrate that the proposed method outperforms the state of the art when only single RGB modality is used for pre-training.

poster session 1 @ red 1, poster session 11 @ red 1, oral session 1 @ red 1, poster session 1, poster session 11, oral session 1

Multimodal sensors (visual, non-visual, and wearable) can provide complementary information to develop robust perception systems for recognizing activities accurately. However, it is challenging to extract robust multimodal representations due to the heterogeneous characteristics of data from multimodal sensors and disparate human activities, especially in the presence of noisy and misaligned sensor data. In this work, we propose a cooperative multitask learning-based guided multimodal fusion approach, MuMu, to extract robust multimodal representations for human activity recognition (HAR). MuMu employs an auxiliary task learning approach to extract features specific to each set of activities with shared characteristics (activity-group). MuMu then utilizes activity-group-specific features to direct our proposed Guided Multimodal Fusion Approach (GM-Fusion) for extracting complementary multimodal representations, designed as the target task. We evaluated MuMu by comparing its performance to state-of-the-art multimodal HAR approaches on three activity datasets. Our extensive experimental results suggest that MuMu outperforms all the evaluated approaches across all three datasets. Additionally, the ablation study suggests that MuMu significantly outperforms the baseline models (p<0.05), which do not use our guided multimodal fusion. Finally, the robust performance of MuMu on noisy and misaligned sensor data posits that our approach is suitable for HAR in real-world settings.

poster session 1 @ red 1, poster session 11 @ red 1, oral session 1 @ red 1, poster session 1, poster session 11, oral session 1

We present Neural Marionette, an unsupervised approach that discovers the skeletal structure from a dynamic sequence and learns to generate diverse motions that are consistent with the observed motion dynamics. Given a video stream of point cloud observation of an articulated body in arbitrary poses, our approach discovers the unknown low-dimensional skeletal relationship that can effectively represent the motion. Then the discovered structure is utilized to encode the motion priors of dynamic sequences in a latent structure, which can be decoded to the relative joint rotations to represent the full skeletal motion. Our approach works without any prior knowledge of the underlying motion or skeletal structure, and we demonstrate that the discovered structure is even comparable to the hand-labeled ground truth skeleton in representing a 4D sequence of motion. The skeletal structure embeds the general semantics of possible motion space that can generate motions for diverse scenarios. We verify that the learned motion prior is generalizable to the multi-modal sequence generation, interpolation of two poses, and motion retargeting to a different skeletal structure.

poster session 1 @ red 1, poster session 11 @ red 1, oral session 1 @ red 1, poster session 1, poster session 11, oral session 1

Semantic scene completion (SSC) aims at predicting the semantic categories and 3D volumetric occupancies of a scene simultaneously. The RGB-D images have been broadly used for providing the semantic and geometry information of the objects in existing semantic scene completion methods. However, existing works use concatenation, element-wise summation and weighted summation to fuse RGB-D data. These strategies ignore the large discrepancy of RGB-D data and the uncertainty measurements of depth data. To solve this problem, we propose the \emph{Frequency Fusion Network} (FFNet), a novel method boosting semantic scene completion by better utilizing RGB-D data. It first explicitly correlates the RGB-D data in the frequency domain, which is different from the features directly extracted by the convolution operation. Then, the correlation information is used to guide the RGB-assisted depth features and depth-assisted RGB features. Further, considering the properties of different frequency components of RGB-D features, we propose a learnable elliptical mask to decompose the features, and attend to different frequency bands to facilitate the correlation process of RGB-D data. We evaluate FFNet intensively on the public SSC benchmarks, where FFNet surpasses the state-of-the-art methods. Our code will be made publicly available.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Weakly supervised Video Anomaly Detection (VAD) using Multi-Instance Learning (MIL) is usually based on the fact that the anomaly score of an abnormal snippet is higher than that of a normal snippet. In the beginning of training, due to the limited accuracy of the model, it is easy to select the wrong abnormal snippet. In order to reduce the probability of selection errors, we first propose a Multi-Sequence Learning (MSL) method and a hinge-based MSL ranking loss that uses a sequence composed of multiple snippets as an optimization unit. We then design a Transformer-based MSL network to learn both video-level anomaly probability and snippet-level anomaly scores. In the inference stage, we propose to use the video-level anomaly probability to suppress the fluctuation of snippet-level anomaly scores. Finally, since VAD needs to predict the snippet-level anomaly scores, by gradually reducing the length of selected sequence, we propose a self-training strategy to gradually refine the anomaly scores. Experimental results show that our method achieves significant improvements on ShanghaiTech, UCF-Crime, and XD-Violence.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Unsupervised video anomaly detection, a task that requires no labeled normal/abnormal training data in any form, is challenging yet of great importance to both industrial applications and academic research. Existing methods typically follow an iterative pseudo label generation process. However, they lack a principled analysis of the impact of such pseudo label generation on training. Furthermore, the long-range temporal dependencies also has been overlooked, which is unreasonable since the definition of an abnormal event depends on the long-range temporal context. To this end, first, we propose a causal graph to analyze the confounding effect of the pseudo label generation process. Then, we introduce a simple yet effective causal inference based framework to disentangle the noisy pseudo label's impact. Finally, we perform counterfactual based model ensemble that blends long-range temporal context with local image context in inference to make final anomaly detection. Extensive experiments on six standard benchmark datasets show that our proposed method significantly outperforms previous state-of-the-art methods, demonstrating our framework's effectiveness.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Video anomaly detection aims to automatically identify unusual objects or behaviours by learning from normal videos. Previous methods tend to use simplistic reconstruction or prediction constraints, which leads to the insufficiency of learned representations for normal data. As such, we propose a novel bi-directional architecture with three consistency constraints to comprehensively regularize the prediction task from pixel-wise, cross-modal, and temporal-sequence levels. First, predictive consistency is proposed to consider the symmetry property of motion and appearance in forwards and backwards time, which ensures the highly realistic appearance and motion predictions at the pixel-wise level. Second, association consistency considers the relevance between different modalities and uses one modality to regularize the prediction of another one. Finally, temporal consistency utilizes the relationship of the video sequence and ensures that the predictive network generates temporally consistent frames. During inference, the pattern of abnormal frames is unpredictable and will therefore cause higher prediction errors. Experiments show that our method outperforms advanced anomaly detectors and achieves state-of-the-art results on UCSD Ped2, CUHK Avenue, and ShanghaiTech datasets.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Although considerable progress has been achieved regarding the transformers in recent years, the large number of parameters, quadratic computational complexity and memory cost conditioned on long sequences make the transformers hard to train and implement, especially in edge computing configurations. In this case, a dizzying number of works have sought to make improvements around computational and memory efficiency upon the original transformer architecture. Nevertheless, many of them restrict the context in the attention to seek a trade-off between cost and performance with prior knowledge of orderly stored data. It is imperative to dig deep into an efficient feature extractor for point clouds due to their irregularity and large quantity of points. In this paper, we propose a novel skeleton decomposition-based self-attention (SD-SA) which has no sequence length limit and exhibits favorable scalability in long-sequence models. Specifically, due to the numerical low-rank nature of self-attention, we approximate it by the skeleton decomposition method while maintaining its effectiveness. At this point, we have shown that the proposed method works for the proposed approach on point cloud classification, segmentation and detection tasks on the ModelNet40, ShapeNet, and KITTI dataset, respectively. Our approach significantly improves the efficiency of the point cloud transformer and exceeds other efficient transformers on point cloud tasks in terms of speed at comparable performance.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Even though considerable progress has been made in deep learning-based 3D point cloud processing, how to obtain accurate correspondences for robust registration remains a major challenge because existing hard assignment methods cannot deal with outliers naturally. Alternatively, the soft matching-based methods have been proposed to learn the matching probability rather than hard assignment. However, in this paper, we prove that these methods have an inherent ambiguity causing many deceptive correspondences. To address the above challenges, we propose to learn a partial permutation matching matrix, which does not assign corresponding points to outliers, and implements hard assignment to prevent ambiguity. However, this proposal poses two new problems, \ie existing hard assignment algorithms can only solve a full rank permutation matrix rather than a partial permutation matrix, and this desired matrix is defined in the discrete space, which is non-differentiable. In response, we design a dedicated soft-to-hard (S2H) matching procedure within the registration pipeline consisting of two steps: solving the soft matching matrix (S-step) and projecting this soft matrix to the partial permutation matrix (H-step). Specifically, we augment the profit matrix before the hard assignment to solve an augmented permutation matrix, which is cropped to achieve the final partial permutation matrix. Moreover, to guarantee end-to-end learning, we supervise the learned partial permutation matrix but propagate the gradient to the soft matrix instead. Our S2H matching procedure can be easily integrated with existing registration frameworks, which has been verified in representative frameworks including DCP, RPMNet, and DGR. Extensive experiments have validated our method, which creates a new state-of-the-art performance for robust 3D point cloud registration. The code will be made public.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Temporal action detection aims to locate the boundaries of action in the video. The current method based on boundary matching enumerates and calculates all possible boundary matchings to generate proposals. However, these methods neglect the long-range context aggregation in boundary prediction. At the same time, due to the similar semantics of adjacent matchings, local semantic aggregation of densely-generated matchings cannot improve semantic richness and discrimination. In this paper, we propose the end-to-end proposal generation method named Dual Context Aggregation Network (DCAN) to aggregate context on two levels, namely, boundary level and proposal level, for generating high-quality action proposals, thereby improving the performance of temporal action detection. Specifically, we design the Multi-Path Temporal Context Aggregation (MTCA) to achieve smooth context aggregation on boundary level and precise evaluation of boundaries. For matching evaluation, Coarse-to-fine Matching (CFM) is designed to aggregate context on the proposal level and refine the matching map from coarse to fine. We conduct extensive experiments on ActivityNet v1.3 and THUMOS-14. DCAN obtains an average mAP of 35.39% on ActivityNet v1.3 and reaches mAP 54.14% at IoU@0.5 on THUMOS-14, which demonstrates DCAN can generate high-quality proposals and achieve state-of-the-art performance. We release the code at https://github.com/cg1177/DCAN.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Learning visual knowledge from massive weakly-labeled web videos has attracted growing research interests thanks to the large corpus of easily accessible video data on the Internet. However, for video action recognition, the action of interest might only exist in arbitrary clips of untrimmed web videos, resulting in high label noises in the temporal space. To address this challenge, we introduce a new method for pre-training video action recognition models using queried web videos. Instead of trying to filter out potential noises, we propose to provide fine-grained supervision signals by defining the concept of Sub-Pseudo Label (SPL). Specifically, SPL spans out a new set of meaningful "middle ground" label space constructed by extrapolating the original weak labels during video querying and the prior knowledge distilled from a teacher model. Consequently, SPL provides enriched supervision for video models to learn better representations and improves data utilization efficiency of untrimmed videos. We validate the effectiveness of our method on four video action recognition datasets and a weakly-labeled image dataset. Experiments show that SPL outperforms several existing pre-training strategies and the learned representations lead to competitive results on several benchmarks.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Previous deep learning-based line segment detection (LSD) suffers from the immense model size and high computational cost for line prediction. This constrains them from real-time inference on computationally restricted environments. In this paper, we propose a real-time and light-weight line segment detector for resource-constrained environments named Mobile LSD (M-LSD). We design an extremely efficient LSD architecture by minimizing the backbone network and removing the typical multi-module process for line prediction found in previous methods. To maintain competitive performance with a light-weight network, we present novel training schemes: Segments of Line segment (SoL) augmentation, matching and geometric loss. SoL augmentation splits a line segment into multiple subparts, which are used to provide auxiliary line data during the training process. Moreover, the matching and geometric loss allow a model to capture additional geometric cues. Compared with TP-LSD-Lite, previously the best real-time LSD method, our model (M-LSD-tiny) achieves competitive performance with 2.5% of model size and an increase of 130.5% in inference speed on GPU. Furthermore, our model runs at 56.8 FPS and 48.6 FPS on the latest Android and iPhone mobile devices, respectively. To the best of our knowledge, this is the first real-time deep LSD available on mobile devices. Our code is available.

poster session 1 @ red 1, poster session 11 @ red 1, oral session 11 @ red 1, poster session 1, poster session 11, oral session 11

Temporal action proposal generation (TAPG) is a challenging task that aims to locate action instances in untrimmed videos with temporal boundaries. To evaluate the confidence of proposals, the existing works typically predict action score of proposals that are supervised by the temporal Intersection-over-Union (tIoU) between proposal and the ground-truth. In this paper, we innovatively propose a general auxiliary \emph{Background Constraint} idea to further suppress low-quality proposals, by utilizing the background prediction score to restrict the confidence of proposals. In this way, the Background Constraint concept can be easily plug-and-played into existing TAPG methods (\emph{e.g.}, BMN, GTAD). From this perspective, we propose the Background Constraint Network (BCNet) to further take advantage of the rich information of action and background. Specifically, we introduce an Action-Background Interaction module for reliable confidence evaluation, which models the inconsistency between action and background by attention mechanisms at the frame and clip levels. Extensive experiments are conducted on two popular benchmarks, \emph{i.e.}, ActivityNet-1.3 and THUMOS14. The results demonstrate that our method outperforms state-of-the-art methods. Equipped with the existing action classifier, our method also achieves remarkable performance on the temporal action localization task.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

In point cloud compression, sufficient contexts are significant for modeling the point cloud distribution. However, the contexts gathered by the previous voxel-based methods decrease when handling sparse point clouds. To address this problem, we propose a multiple-contexts deep learning framework called OctAttention employing the octree structure, a memory-efficient representation for point clouds. Our approach encodes octree symbol sequences in a lossless way by gathering the information of sibling and ancestor nodes. Expressly, we first represent point clouds with octree to reduce spatial redundancy, which is robust for point clouds with different resolutions. We then design a conditional entropy model with a large receptive field that models the sibling and ancestor contexts to exploit the strong dependency among the neighboring nodes and employ an attention mechanism to emphasize the correlated nodes in the context. Furthermore, we introduce a mask operation during training and testing to make a trade-off between encoding time and performance. Compared to the previous state-of-the-art works, our approach obtains a 10%-35% BD-Rate gain on the Lidar benchmark (e.g. SemanticKITTI) and object point cloud dataset (e.g. MPEG 8i, MVUB), and saves 95% coding time compared to the voxel-based baseline. The code is available at https://github.com/zb12138/OctAttention.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Pedestrian trajectory prediction is crucial in many practical applications due to the diversity of pedestrian movements, such as social interactions and individual motion behaviors. With similar observable trajectories and social environments, different pedestrians may make completely different future decisions. However, most existing methods only focus on the frequent modal of the trajectory and thus are difficult to generalize to the peculiar scenario, which leads to the decline of the multimodal fitting ability when facing similar scenarios. In this paper, we propose a complementary attention gated network (CAGN) for pedestrian trajectory prediction, in which a dual-path architecture including normal and inverse attention is proposed to capture both frequent and peculiar modals in spatial interactions and temporal motion patterns, respectively. Specifically, a complementary block is proposed to guide normal and inverse attention, which are then be summed with learnable weights to get attention features by a gated network. Subsequently, multiple trajectory distributions are estimated based on the fused spatio-temporal attention features due to the multimodality of future trajectory. Experimental results on the popular benchmark datasets, i.e., the ETH, and the UCY, demonstrate that our method outperforms state-of-the-art methods by 13.8% in Average Displacement Error (ADE) and 10.4% in Final Displacement Error (FDE).

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Temporal action segmentation classifies the action of each frame in (long) video sequences. Due to the high cost of frame-wise labeling, we propose the first semi-supervised method for temporal action segmentation. Our method hinges on unsupervised representation learning, which, for temporal action segmentation, poses unique challenges. Actions in untrimmed videos vary in length and have unknown labels and start/end times. Ordering of actions across videos may also vary. We propose a novel way to learn frame-wise representations from temporal convolutional networks (TCNs) by clustering input features with added time-proximity condition and multi-resolution similarity. By merging representation learning with conventional supervised learning, we develop an ``Iterative-Contrast-Classify (ICC)'' semi-supervised learning scheme. With more labelled data, ICC progressively improves in performance; ICC semi-supervised learning, with 40% labelled videos, performs similar to fully-supervised counterparts. Our ICC improves MoF by {+1.8, +5.6, +2.5}% on Breakfast, 50Salads and GTEA respectively for 100% labelled videos.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

This paper proposes a novel approach to object detection on drone imagery, namely Multi-Proxy Detection Network with Unified Foreground Packing (UFPMP-Det). To deal with the numerous instances of very small scales, different from the common solution that divides the high-resolution input image into quite a number of chips with low foreground ratios to perform detection on them each, the Unified Foreground Packing (UFP) module is designed, where the sub-regions given by a coarse detector are initially merged through clustering to suppress background and the resulting ones are subsequently packed into a mosaic for a single inference, thus significantly reducing overall time cost. Furthermore, to address the more serious confusion between inter-class similarities and intra-class variations of instances, which deteriorates detection performance but is rarely discussed, the Multi-Proxy Detection Network (MP-Det) is presented to model object distributions in a fine-grained manner by employing multiple proxy learning, and the proxies are enforced to be diverse by minimizing a Bag-of-Instance-Words (BoIW) guided optimal transport loss. By such means, UFPMP-Det largely promotes both the detection accuracy and efficiency. Extensive experiments are carried out on the widely used VisDrone and UAVDT datasets, and UFPMP-Det reports new state-of-the-art scores at a much higher speed, highlighting its advantages.

poster session 1 @ red 1, poster session 11 @ red 1, oral session 11 @ red 1, poster session 1, poster session 11, oral session 11

In this paper, we present a new self-supervised scene flow estimation approach for a pair of consecutive point clouds. The key idea of our approach is to represent discrete point clouds as continuous probability density functions using Gaussian mixture models. Scene flow estimation is therefore converted into the problem of recovering motion from the alignment of probability density functions, which we achieve using a closed-form expression of the classic Cauchy-Schwarz divergence. Unlike existing nearest-neighbor-based approaches that use hard pairwise correspondences, our proposed approach establishes soft and implicit point correspondences between point clouds and generates more robust and accurate scene flow in the presence of missing correspondences and outliers. Comprehensive experiments show that our method makes noticeable gains over the Chamfer Distance and the Earth Mover’s Distance in real-world environments and achieves state-of-the-art performance among self-supervised learning methods on FlyingThings3D and KITTI, even outperforming some supervised methods with ground truth annotations.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

We present a novel non-iterative learnable method for partial-to-partial 3D shape registration. The partial alignment task is extremely complex, as it jointly tries to match between points, and identify which points do not appear in the corresponding shape, causing the solution to be non-unique and ill-posed in most cases. Until now, two main methodologies have been suggested to solve this problem: sample a subset of points that are likely to have correspondences, or perform soft alignment between the point clouds and try to avoid a match to an occluded part. These heuristics work when the partiality is mild or when the transformation is small but fails for severe occlusions, or when outliers are present. We present a novel approach named Confidence Guided Distance Network, where we fuse learnable similarity between point embeddings and spatial distance between point clouds, inducing an optimized solution for the overlapping points while ignoring parts that only appear in one of the shapes. The point feature generation is done by a self-supervised architecture that repels far points to have different embeddings, therefore succeeds to align partial views of shapes, even with excessive internal symmetries, or acute rotations. We compare our network to recently presented learning-based and axiomatic methods and report a fundamental boost in performance.

poster session 1 @ red 1, poster session 11 @ red 1, oral session 11 @ red 1, poster session 1, poster session 11, oral session 11

Voxel and point representations are widely applied in recent 3D object detection tasks from LiDAR point clouds. Voxel representations contribute to efficiently and rapidly locating objects, whereas point representations are capable of describing intra-object spatial relationship for detection refinement. In this work, we aim to exploit the strengths of both two representations, and present a novel two-stage detector, named Joint Point-Voxel Network (JPV-Net). Specifically, our framework is equipped with a Dual Encoders-Fusion Decoder, which consists of the dual encoders to extract voxel features of sketchy 3D scenes and point features rich in geometric context, respectively, and the Feature Propagation Fusion (FP-Fusion) decoder to attentively fuse them from coarse to fine. By making use of the advantages of these features, the refinement network can effectively eliminate false detection and provide better accuracy. Besides, to further develop the perception characteristics of voxel CNN and point backbone, we design two novel intersection-over-union (IoU) estimation modules for proposal generation and refinement, both of which can alleviate the misalignment between the localization and the classification confidence. Extensive experiments on the KITTI dataset and ONCE dataset demonstrate that our proposed JPV-Net outperforms other state-of-the-art methods with remarkable margins.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

A common approach to Temporal Action Localization (TAL) is to generate action proposals and then perform action classification and localization on them. For each proposal, existing methods universally use a shared proposal-level representation for both tasks. However, our analysis indicates that this shared representation focuses on the most discriminative frames for classification, e.g., ``take-offs" rather than ``run-ups" in distinguishing ``high jump" and ``long jump", while frames most relevant to localization, such as the start and end frames of an action, are largely ignored. In other words, such a shared representation can not simultaneously handle both classification and localization tasks well, and it makes precise TAL difficult. To address this challenge, this paper disentangles the shared representation into classification and localization representations. The disentangled classification representation focuses on the most discriminative frames, and the disentangled localization representation focuses on the action phase as well as the action start and end. Our model could be divided into two sub-networks, i.e., the disentanglement network and the context-based aggregation network. The disentanglement network is an autoencoder to learn orthogonal hidden variables of classification and localization. The context-based aggregation network aggregates the classification and localization representations by modeling local and global contexts. We evaluate our proposed method on two popular benchmarks for TAL, which outperforms all state-of-the-art methods.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11

Weakly-supervised temporal action localization (WTAL) in untrimmed videos has emerged as a practical but challenging task since only video-level labels are available. Existing approaches typically leverage off-the-shelf segment-level features, which suffer from spatial incompleteness and temporal incoherence, thus limiting their performance. In this paper, we tackle this problem from a new perspective by enhancing segment-level representations with a simple yet effective graph convolutional network, namely action complement graph network (ACGNet). It facilitates the current video segment to perceive spatial-temporal dependencies from others that potentially convey complementary clues, implicitly mitigating the negative effects caused by the two issues above. By this means, the segment-level features are more discriminative and robust to spatial-temporal variations, contributing to higher localization accuracies. More importantly, the proposed ACGNet works as a universal module that can be flexibly plugged into different WTAL frameworks, while maintaining the end-to-end training fashion. Extensive experiments are conducted on the THUMOS'14 and ActivityNet1.2 benchmarks, where the state-of-the-art results clearly demonstrate the superiority of the proposed approach.

poster session 1 @ red 1, poster session 11 @ red 1, poster session 1, poster session 11