Advanced 3D Object Detection with Tone-Mapping Enhanced Fine Disparity R-CNN

Abstract

3D object detection is one of the key techniques in autonomous driving. However, existing 3D object detectors will likely be impaired under low illumination conditions and show poor performance on small object detection. To resolve this limitation, this paper proposes a tone mapping-based fine disparity R-CNN (TFD R-CNN) for effective 3D object detection. We first develop a tone mapping technique to enhance the color contrasts and thus to make objects clearer, for easier detection of low illuminated objects. Furthermore, optimization of grouping radius has been done to improve the detection performance on small-sized objects. By reducing the grouping radius, the features of small objects can be preserved, so that more refined feature information can be extracted. Experimental result comparisons show that the proposed TFD R-CNN outperforms the well-known Disp R-CNN. Specifically, the proposed method can achieve about 4% better performance in average precision than the baseline Disp R-CNN on the KITTI dataset.