报告题目：MonoBins: Self-Supervised Monocular Depth Estimation Based on Depth Bins

报告人：方效林副教授（东南大学计算机科学与技术学院）

报告时间：2022年11月25日（周五）上午10:00-11:00

报告地点（线上）：腾讯会议ID：867 973 010

摘要：Current regression based depth estimation methods from monocular RGB images face with the difficulty that the depth distributions are quite different in diverse scenes. Therefore, we propose a classification based self-supervised depth estimation approach by exploiting the idea of dividing the depth range into multiple bins. After calculating the probabilities that the depth of a pixel locates in each bin, it can obtain the final depth of that pixel by a combination of the probabilities and the bins' centers. We call our framework of depth estimation as MonoBins. We also consider the influence of the moving objects by encoding geometric priori knowledge from adjacent frames into a Cost Volume structure, and a moving object mask is designed to eliminate their interference for the training process. We validate the effectiveness of the proposed method on the KITTI dataset, and the results show its significant performance improvement on most of the metrics. In particular, our method can reduce the RMSE and S*Rel errors by 17.8% and 16.5% for low resolution images, and 25.9% and 17.4% for high resolution images respectively.

邀请人：贾日恒