研究目的
Investigating the effectiveness of asymmetric two-stream networks for RGB-Disparity based object detection to handle hard samples that monocular-based methods struggle with.
研究成果
The asymmetric two-stream networks proposed in this paper effectively combine RGB and disparity data for object detection, achieving significant performance improvements over monocular-based methods. The method is computationally efficient and feasible for public applications due to its reliance on only binocular information.
研究不足
The study is limited to 2D object detection and requires a pair of normal RGB cameras for disparity map generation. The method's performance is dependent on the quality of the disparity maps.
1：Experimental Design and Method Selection:
The study employs asymmetric two-stream networks consisting of Disparity Representations Mining Network (DRMN) and Muti-Modal Detection Network (MMDN) to combine RGB and disparity data for object detection.
2：Sample Selection and Data Sources:
The experiments are performed on the KITTI Object Detection Benchmark and the proposed Binocular Pedestrian Detection (BPD) dataset.
3：List of Experimental Equipment and Materials:
The study uses a pair of normal RGB cameras for capturing images and generating disparity maps.
4：Experimental Procedures and Operational Workflow:
The methodology involves training DRMN and MMDN separately before finetuning the combined networks.
5：Data Analysis Methods:
The performance is evaluated using mAP with IoU of 0.5 as the criteria, and speed is measured with batch size 1 using TITAN X with Intel Xeon E5-2620@2.10 GHz.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
