Improved Pseudo-Stereo for Monocular 3D Object Detection in Autonomous Driving

Monocular 3D object detection is a challenging task due to the lack of accurate 3D information in single images. Stereo 3D detectors, which can accurately localize 3D objects, have made significant progress. Inspired by this, we propose a “Multi-Scale Attention Fusion” (MSFA) to enhance the PseudoStereo method. First, a more accurate depth map than the baseline method is obtained from the input left image using a pre-trained depth estimation network, FeatDepth. Then, a MSFA embeds disparity features into the left image features to generate complementary disparity features. Finally, the effectiveness of the method is verified through ablation experiments. The experimental results show that compared to the baseline algorithm on the KITTI dataset, our algorithm improves bird’seye view object detection task performance by 1.22%, 0.59%, and 0.14% for easy, moderate, and hard settings, respectively.