August 2020
tl;dr: Mono for stereo. Learn stereo matching with monocular images.
The basic idea is to generate stereo training pair with mono depth to train stereo matching algorithms. This idea is very similar to that of Homographic Adaptation in SuperPoint, in that both generates training data and GT with known geometric transformation.
This still need a stereo pair as input during inference time. The main idea is to use monodepth to predict a depth map, sharpen it, and generate a stereo pair, with known stereo matching GT.
- Summaries of the key ideas
- Training uses PSMNet (pyramid stereo matching)