Stereo 3D Object Detection Using a Feature Attention Module

Stereo 3D object detection remains a crucial challenge within the realm of 3D vision. In the pursuit of enhancing stereo 3D object detection, feature fusion has emerged as a potent strategy. However, the design of the feature fusion module and the determination of pivotal features in this fusion process remain critical. This paper proposes a novel feature attention module tailored for stereo 3D object detection. Serving as a pivotal element for feature fusion, this module not only discerns feature importance but also facilitates informed enhancements based on its conclusions. This study delved into the various facets aided by the feature attention module. Firstly, a interpretability analysis was conducted concerning the function of the image segmentation methods. Secondly, we explored the augmentation of the feature fusion module through a category reweighting strategy. Lastly, we investigated global feature fusion methods and model compression strategies. The models devised through our proposed design underwent an effective analysis, yielding commendable performance, especially in small object detection within the pedestrian category.