Wenxiao Wang

Chenxi Huang, Tong He, Haidong Ren et al.

Compared to typical multi-sensor systems, monocular 3D object detection has attracted much attention due to its simple configuration. However, there is still a significant gap between LiDAR-based and monocular-based methods. In this paper, we find that the ill-posed nature of monocular imagery can lead to depth ambiguity. Specifically, objects with different depths can appear with the same bounding boxes and similar visual features in the 2D image. Unfortunately, the network cannot accurately distinguish different depths from such non-discriminative visual features, resulting in unstable depth training. To facilitate depth learning, we propose a simple yet effective plug-and-play module, \underline{O}ne \underline{B}ounding Box \underline{M}ultiple \underline{O}bjects (OBMO). Concretely, we add a set of suitable pseudo labels by shifting the 3D bounding box along the viewing frustum. To constrain the pseudo-3D labels to be reasonable, we carefully design two label scoring strategies to represent their quality. In contrast to the original hard depth labels, such soft pseudo labels with quality scores allow the network to learn a reasonable depth range, boosting training stability and thus improving final performance. Extensive experiments on KITTI and Waymo benchmarks show that our method significantly improves state-of-the-art monocular 3D detectors by a significant margin (The improvements under the moderate setting on KITTI validation set are $\mathbf{1.82\sim 10.91\%}$ \textbf{mAP in BEV} and $\mathbf{1.18\sim 9.36\%}$ \textbf{mAP in 3D}). Codes have been released at \url{https://github.com/mrsempress/OBMO}.

Shuai Zhao, Liguang Zhou, Wenxiao Wang et al.

The width of a neural network matters since increasing the width will necessarily increase the model capacity. However, the performance of a network does not improve linearly with the width and soon gets saturated. In this case, we argue that increasing the number of networks (ensemble) can achieve better accuracy-efficiency trade-offs than purely increasing the width. To prove it, one large network is divided into several small ones regarding its parameters and regularization components. Each of these small networks has a fraction of the original one's parameters. We then train these small networks together and make them see various views of the same data to increase their diversity. During this co-training process, networks can also learn from each other. As a result, small networks can achieve better ensemble performance than the large one with few or no extra parameters or FLOPs, \ie, achieving better accuracy-efficiency trade-offs. Small networks can also achieve faster inference speed than the large one by concurrent running. All of the above shows that the number of networks is a new dimension of model scaling. We validate our argument with 8 different neural architectures on common benchmarks through extensive experiments. The code is available at \url{https://github.com/FreeformRobotics/Divide-and-Co-training}.

Ping Li, Chenhao Ping, Wenxiao Wang et al.

Knowledge Distillation (KD) compresses neural networks by learning a small network (student) via transferring knowledge from a pre-trained large network (teacher). Many endeavours have been devoted to the image domain, while few works focus on video analysis which desires training much larger model making it be hardly deployed in resource-limited devices. However, traditional methods neglect two important problems, i.e., 1) Since the capacity gap between the teacher and the student exists, some knowledge w.r.t. difficult-to-transfer samples cannot be correctly transferred, or even badly affects the final performance of student, and 2) As training progresses, difficult-to-transfer samples may become easier to learn, and vice versa. To alleviate the two problems, we propose a Sample-level Adaptive Knowledge Distillation (SAKD) framework for action recognition. In particular, it mainly consists of the sample distillation difficulty evaluation module and the sample adaptive distillation module. The former applies the temporal interruption to frames, i.e., randomly dropout or shuffle the frames during training, which increases the learning difficulty of samples during distillation, so as to better discriminate their distillation difficulty. The latter module adaptively adjusts distillation ratio at sample level, such that KD loss dominates the training with easy-to-transfer samples while vanilla loss dominates that with difficult-to-transfer samples. More importantly, we only select those samples with both low distillation difficulty and high diversity to train the student model for reducing computational cost. Experimental results on two video benchmarks and one image benchmark demonstrate the superiority of the proposed method by striking a good balance between performance and efficiency.