Predicting attention regions of interest is an important yet challenging task for self-driving systems. Existing methodologies rely on large-scale labeled traffic datasets that are labor-intensive to obtain. Besides, the huge domain gap between natural scenes and traffic scenes in current datasets also limits the potential for model training. To address these challenges, we are the first to introduce an unsupervised way to predict self-driving attention by uncertainty modeling and driving knowledge integration. Our approach's Uncertainty Mining Branch (UMB) discovers commonalities and differences from multiple generated pseudo-labels achieved from models pre-trained on natural scenes by actively measuring the uncertainty. Meanwhile, our Knowledge Embedding Block (KEB) bridges the domain gap by incorporating driving knowledge to adaptively refine the generated pseudo-labels. Quantitative and qualitative results with equivalent or even more impressive performance compared to fully-supervised state-of-the-art approaches across all three public datasets demonstrate the effectiveness of the proposed method and the potential of this direction. The code will be made publicly available.
Weakly-supervised temporal action localization aims to locate action regions and identify action categories in untrimmed videos, only taking video-level labels as the supervised information. Pseudo label generation is a promising strategy to solve the challenging problem, but most existing methods are limited to employing snippet-wise classification results to guide the generation, and they ignore that the natural temporal structure of the video can also provide rich information to assist such a generation process. In this paper, we propose a novel weakly-supervised temporal action localization method by inferring snippet-feature affinity. First, we design an affinity inference module that exploits the affinity relationship between temporal neighbor snippets to generate initial coarse pseudo labels. Then, we introduce an information interaction module that refines the coarse labels by enhancing the discriminative nature of snippet-features through exploring intra- and inter-video relationships. Finally, the high-fidelity pseudo labels generated from the information interaction module are used to supervise the training of the action localization network. Extensive experiments on two publicly available datasets, i.e., THUMOS14 and ActivityNet v1.3, demonstrate our proposed method achieves significant improvements compared to the state-of-the-art methods.
In this paper, we propose the semantic graph Transformer (SGT) for the 3D scene graph generation. The task aims to parse a cloud point-based scene into a semantic structural graph, with the core challenge of modeling the complex global structure. Existing methods based on graph convolutional networks (GCNs) suffer from the over-smoothing dilemma and could only propagate information from limited neighboring nodes. In contrast, our SGT uses Transformer layers as the base building block to allow global information passing, with two types of proposed Transformer layers tailored for the 3D scene graph generation task. Specifically, we introduce the graph embedding layer to best utilize the global information in graph edges while maintaining comparable computation costs. Additionally, we propose the semantic injection layer to leverage categorical text labels and visual object knowledge. We benchmark our SGT on the established 3DSSG benchmark and achieve a 35.9% absolute improvement in relationship prediction's R@50 and an 80.40% boost on the subset with complex scenes over the state-of-the-art. Our analyses further show SGT's superiority in the long-tailed and zero-shot scenarios. We will release the code and model.
Video denoising aims to recover high-quality frames from the noisy video. While most existing approaches adopt convolutional neural networks(CNNs) to separate the noise from the original visual content, however, CNNs focus on local information and ignore the interactions between long-range regions. Furthermore, most related works directly take the output after spatio-temporal denoising as the final result, neglecting the fine-grained denoising process. In this paper, we propose a Dual-stage Spatial-Channel Transformer (DSCT) for coarse-to-fine video denoising, which inherits the advantages of both Transformer and CNNs. Specifically, DSCT is proposed based on a progressive dual-stage architecture, namely a coarse-level and a fine-level to extract dynamic feature and static feature, respectively. At both stages, a Spatial-Channel Encoding Module(SCEM) is designed to model the long-range contextual dependencies at spatial and channel levels. Meanwhile, we design a Multi-scale Residual Structure to preserve multiple aspects of information at different stages, which contains a Temporal Features Aggregation Module(TFAM) to summarize the dynamic representation. Extensive experiments on four publicly available datasets demonstrate our proposed DSCT achieves significant improvements compared to the state-of-the-art methods.
The ever-increasing heavy traffic congestion potentially impedes the accessibility of emergency vehicles (EVs), resulting in detrimental impacts on critical services and even safety of people's lives. Hence, it is significant to propose an efficient scheduling approach to help EVs arrive faster. Existing vehicle-centric scheduling approaches aim to recommend the optimal paths for EVs based on the current traffic status while the road-centric scheduling approaches aim to improve the traffic condition and assign a higher priority for EVs to pass an intersection. With the intuition that real-time vehicle-road information interaction and strategy coordination can bring more benefits, we propose LEVID, a LEarning-based cooperative VehIcle-roaD scheduling approach including a real-time route planning module and a collaborative traffic signal control module, which interact with each other and make decisions iteratively. The real-time route planning module adapts the artificial potential field method to address the real-time changes of traffic signals and avoid falling into a local optimum. The collaborative traffic signal control module leverages a graph attention reinforcement learning framework to extract the latent features of different intersections and abstract their interplay to learn cooperative policies. Extensive experiments based on multiple real-world datasets show that our approach outperforms the state-of-the-art baselines.
For a new city that is committed to promoting Electric Vehicles (EVs), it is significant to plan the public charging infrastructure where charging demands are high. However, it is difficult to predict charging demands before the actual deployment of EV chargers for lack of operational data, resulting in a deadlock. A direct idea is to leverage the urban transfer learning paradigm to learn the knowledge from a source city, then exploit it to predict charging demands, and meanwhile determine locations and amounts of slow/fast chargers for charging stations in the target city. However, the demand prediction and charger planning depend on each other, and it is required to re-train the prediction model to eliminate the negative transfer between cities for each varied charger plan, leading to the unacceptable time complexity. To this end, we propose the concept and an effective solution of Simultaneous Demand Prediction And Planning (SPAP): discriminative features are extracted from multi-source data, and fed into an Attention-based Spatial-Temporal City Domain Adaptation Network (AST-CDAN) for cross-city demand prediction; a novel Transfer Iterative Optimization (TIO) algorithm is designed for charger planning by iteratively utilizing AST-CDAN and a charger plan fine-tuning algorithm. Extensive experiments on real-world datasets collected from three cities in China validate the effectiveness and efficiency of SPAP. Specially, SPAP improves at most 72.5% revenue compared with the real-world charger deployment.
We address the challenging task of cross-modal moment retrieval, which aims to localize a temporal segment from an untrimmed video described by a natural language query. It poses great challenges over the proper semantic alignment between vision and linguistic domains. Most of these methods only leverage sentences in the multi-modal fusion stage and independently extract the features of videos and sentences, which do not make full use of the potential of language. In this paper, we present Language Guided Networks (LGN), a new framework that tightly integrates cross-modal features in multiple stages. In the first feature extraction stage, we introduce to capture the discriminative visual features which can cover the complex semantics in the sentence query. Specifically, the early modulation unit is designed to modulate convolutional feature maps by a linguistic embedding. Then we adopt a multi-modal fusion module in the second fusion stage. Finally, to get a precise localizer, the sentence information is utilized to guide the process of predicting temporal positions. Specifically, the late guidance module is developed to further bridge vision and language domain via the channel attention mechanism. We evaluate the proposed model on two popular public datasets: Charades-STA and TACoS. The experimental results demonstrate the superior performance of our proposed modules on moment retrieval (improving 5.8\% in terms of R1@IoU5 on Charades-STA and 5.2\% on TACoS). We put the codes in the supplementary material and will make it publicly available.
Deep neural networks have achieved remarkable success for video-based action recognition. However, most of existing approaches cannot be deployed in practice due to the high computational cost. To address this challenge, we propose a new real-time convolutional architecture, called Temporal Convolutional 3D Network (T-C3D), for action representation. T-C3D learns video action representations in a hierarchical multi-granularity manner while obtaining a high process speed. Specifically, we propose a residual 3D Convolutional Neural Network (CNN) to capture complementary information on the appearance of a single frame and the motion between consecutive frames. Based on this CNN, we develop a new temporal encoding method to explore the temporal dynamics of the whole video. Furthermore, we integrate deep compression techniques with T-C3D to further accelerate the deployment of models via reducing the size of the model. By these means, heavy calculations can be avoided when doing the inference, which enables the method to deal with videos beyond real-time speed while keeping promising performance. Our method achieves clear improvements on UCF101 action recognition benchmark against state-of-the-art real-time methods by 5.4% in terms of accuracy and 2 times faster in terms of inference speed with a less than 5MB storage model. We validate our approach by studying its action representation performance on four different benchmarks over three different tasks. Extensive experiments demonstrate comparable recognition performance to the state-of-the-art methods. The source code and the pre-trained models are publicly available at https://github.com/tc3d.
Recent studies on automatic neural architectures search have demonstrated significant performance, competitive to or even better than hand-crafted neural architectures. However, most of the existing network architecture tend to use residual, parallel structures and concatenation block between shallow and deep features to construct a large network. This requires large amounts of memory for storing both weights and feature maps. This is challenging for mobile and embedded devices since they may not have enough memory to perform inference with the designed large network model. To close this gap, we propose MemNet, an augment-trim learning-based neural network search framework that optimizes not only performance but also memory requirement. Specifically, it employs memory consumption based ranking score which forces an upper bound on memory consumption for navigating the search process. Experiment results show that, as compared to the state-of-the-art efficient designing methods, MemNet can find an architecture which can achieve competitive accuracy and save an average of 24.17% on the total memory needed.