[Contents]

The original text is:

Depth ordered regression network for monocular depth estimation

Huanfu 1 Mingminggong 2,3 Chaohui Wang 4 Kaihan Bartmang herrlich 2 Dacheng Tao 1

Three failures of previous work (especially using deep neural network: DCNN) method:

The idea behind it is to classify the distant ones with coarse granularity.

Then, on the basis of "the classification granularity is coarse when the SID is far away", the regression problem can be transformed into a classification problem.

Discrete continuous distances become some distance intervals.

Divided into three modules

Including:

Full image encoder

Hole convolution

This is where SID is used.

Define our unique loss function:

Iterative optimization algorithm, back propagation, we can finally get an ordered label classification, each category is a distance, such as: 1m, 1. 1m, 1.2m, 1.4m, 2m,/kloc-.

Using: ordered label classification, we can infer the distance.

Use the following formula:

D-tip is the assumed depth/distance.

L is a learned label.

In the original paper of 18, there is such data:

In the kitti ranking of 2 1 year, it is very high.

In fact, according to absRel, Donne ranked first.