LAUSR

Depth map plays an important role in our daily life including automatic driving, intelligent robots, and commercial manufacture. However, existing depth estimation methods are still facing some difficulties in real-life applications. Firstly, using merely point-wise metrics during training usually causes surface distortion due to the lack of geometric restrictions. To tackle this problem, we introduce a normal restriction module to the network to improve the method performance with regards to surface integrity. Secondly, although 3D methods usually achieve higher accuracy by directly concatenating features as cost aggregation, they sacrifice the time and memory efficiency. To make the disparity estimation network light enough for adding the normal restriction, we propose a 2D aggregation method to replace the 3D ones. Experiments prove the effectiveness of our method and we show that our results are competitive to 3D methods under popular benchmarks including KITTI, SceneFlow and Midddlebury.