MVDCANet: An End-to-End Self-Attention-Based Multiview-Dualchannel 3D Object Detection

In recent years, the low resolution of point clouds generated by lidar leads to poor recognition of small objects, e.g., Pedestrians. Therefore, it is becoming essential to improve lidar in recognition of small objects. In this paper, we propose a 3D small object detection method based on multi-view, dual-channel feature extraction, which pays more attention to the multi-view features of the points in the captured point cloud. To discriminate the captured multi-view features, we extract multi-view features of the current point by explicitly using the multi-view feature extraction (MVFE) module. It is a proven method for the more original features from the current point, such as spatial and edges, can be retained. Then, the dual-channel attention feature extraction (DCAFE) module enhances the multi-view features of the current point by adding the dual-channel self-attention mechanism and maps into pseudo-image. Finally, the pseudo-images are fed into the 2D Backbone, which contains the DO-Conv module so that the pseudo-image can train features at different scales. Experimental results on the KITTI object benchmark show that our method can implement the detection algorithm with end-to-end real-time inference speed and improve 6.72 3D mAP and 3.23 3D mAP over existing algorithms in pedestrian and cyclist classes. We demonstrate that the proposed method also learns the point cloud feature encoding in the ablation study. Meanwhile, the architecture of 2D convolution detection provides an accurate and efficient detector for 3D scenes, especially for small object recognition.