Learning-AI

DC: Depth Coefficients for Depth Completion

December 2019

tl;dr: Encode depth in a simplified one-hot encoding (DC) and cross entropy loss reduces over-smoothing in depth estimation.

Overall impression

Similar to the idea of SMWA to address the “long tail” problem. This problem can be also referred to as edge bleeding, over-smoothing, or mixed depth. It features spurious depth estimation in mid-air and connecting surfaces between separate objects.

DC focuses on depth completion while SMWA focuses on depth estimation from stereo pairs.

It also acknowledges that the problem is a multi-modal problem and using L1 or L2 leads to spurious estimation in-between modes. –> this is also used in generalized focal loss to model multi-modal distribution of edges of heavily occluded objects.

The idea of using an N-channel but 3-hot depth encoding is similar to the soft one-hot encoding used in SMWA. In SMWA it also uses cross entropy for regression. DC gives a better explanation why cross entropy is a better loss than L1 or L2.

The input and loss modification is based on sparse-to-dense and is easy to implement.

Key ideas

Technical details

Notes