【Deep Learning】Review of Stereo Matching by Training a Convolutional Neural Network to Compare Image

Stereo Matching by Training a Convolutional Neural Network to Compare Image

Link: http://arxiv.org/abs/1510.05970

Code: https://github.com/jzbontar/mc-cnn

_____________________________________

1.      Summary of thePaper

By presenting a matching cost computation, the paperprovides a similarity measurement extracting depth information from a pairpictures via convolutional neural network. The speed model could give resultswithin a second while the accuracy model lower the error rate on Kitti.

 

2.      MainContributions

1)      Asthe paper itself indicated, a description of two architectures (One for speedand the other for accuracy.) based on convolutional neural networks forcomputing the stereo matching cost.

2)      Amethod, accompanied by its source code, with the lowest error rate on the KITTI2012, KITTI 2015, and Middlebury stereo data sets.

3)      Experimentsanalyzing the importance of data set size, the error rate compared with othermethods, and the trade-off between accuracy and runtime for different settingsof the hyper- parameters.

 

3.      Positive andnegative points

Positive Points:

(i)            Itprovides two models that either gave a fastest speed ever or gave the mostaccurate rate ever since.

(ii)          Itdemonstrates very detailed dataset augmentation methods, which, not novelthough, are necessary to achieve good results.

Negative Points:

(i) It actuallydidn’t present brand-new idea of improving the stereo matching but indeedprovides us an overall perspective of the whole process.

 

4.      How strong isthe evaluation

Extremelyimpressive. The accurate MC-CNN-acrt model ranks first amony all method on theKITTI 2012. Even the MC-CNN-fst model ranks 5^th and the runtime islargely beyond any other methods.

As for KITTIstereo ranking, these two methods totally dominate others as they rank 1^stand 2^nd lowest error rate. However, ELAS methods is slightly quickerthan MC-CNN-fst.

However, theMC-CNN-fst cannot show up in the Middlebury stereo data set but stillMC-CNN-acrt ranks 1^st without doubt.

 

5.      Possibledirection for the future work

Maybe they canpursue the measurement of similarity for a video.