【Deep Learning】Review of Designing Deep Networks for Surface Normal Estimation

DesigningDeep Networks for Surface Normal Estimation
link：http://www.cs.cmu.edu/~dfouhey/deep3d/deep3d.pdf
________________________________________________

1. Summary of thePaper

By injecting insight in 3D representation, the authorsdesigned a new CNN architecture for the better demonstration of surface normalestimation. They achieve this by developing CNNs that operate both locally andglobally and moreover use fusion
method to handle these two forms of evidence.This method obtained the state-of-art performance in surface normal estimation.

2. MainContributions

1) Surprisinglyfusing global and local features and therefore better represent the surface ofobject. It’s beautifully reconciling evidence from conflicting sources.

2) Welldetecting man-made nature of indoor scenes.

3) Takingadvantage of the local constraints, they incorporate them in learning of thelocal network and as an input for the fusion network, which improve theperformance over the simple feed-forward network.

3. Positive andnegative points

Positive Points:

(i) Again,their fusion network that enables global and local constrains to be consideredtogether.

(ii) Basedon that, this method could recognize detail part of furniture in the room.

Negative Points:

(i) Cannot thinkof that.

4. How strong isthe evaluation

Very impressive.According to the Table 1., their network’s mean and median ranked lowest amongNYUv2 for per-pixel surface normal estimation.

5. Possible directionfor the future work

Myguess will be advanced fusion network that integrates the global and localinformation better. Based on this, p