Learning-Deep-Learning

PackNet: 3D Packing for Self-Supervised Monocular Depth Estimation

May 2020

tl;dr: Improvement of neural net architecture to improve self-superv0sed depth. It leverage velocity when possible to achieve scale awareness.

Overall impression

The paper inherits all previous loss functions in self-supervised monocular depth method, and can be seen as a bag-of-trick for self-supervised mono depth.

The importance of high resolution has been demonstrated through SuperDepth ICRA 2019 via super-resolution with channel2spatial.

Most previous method requires depth scaling at test time. When velocity is available for regularization (during training), the network outputs metric results at minor degradation of performance. However it does not leverage this info during test time, and may fail at unseen ego motion. See SC-Sfmlearner and TrianFLow for other scale awareness solution.

Key ideas

Technical details

Notes