Deep Learning paper Notes (3): Deep learning face Attributes in the Wild

This is a technical report by Xiaogang Wang and Xiaoou Tang Group, the author of which is an excellent learning brother Ziwei Liu.

Usually face recognition in the first to detect and align the face image, and then in the corresponding location to extract features, but in the natural scene, due to background confusion, face detection and alignment will be affected, and then affect feature extraction and the final recognition effect.

The main idea of this paper is to build a system of face attributes recognition by learning two deep network, the first of which is used to localization and the second is to extract feature.

Main process:

(a) (b) is Corse to fine the position of the localize face, after (a) CNN gets a picture of the face and shoulder part, after (b) The CNN further gets the picture of the human face part.

(c) CNN is used to learn the features of the human face part of the picture, and finally put the features of fully connected layer together, using (d) SVM to get the classification results.

Learning algorithm:

1. Use the object categories inside the imagenet to go to the Pre-train (a) and (b) of the two CNN. Use to determine whether the picture contains object's task to pre-train a hope to automatically locate the face location of the CNN, there is a certain reason, know object, basically also know the face in which piece.

2. Use the face identity of the celebfaces dataset to go to CNN in Pre-train (c) to extract facial features. As can be seen from (c), it is a little different from the general CNN, in the middle two is a partial convolution, the front and back two layer is the global convolution, the effect of local convolution is to maintain the location of the picture information, specific details of the original paper.

3. Finally, the attribute label is used to fine-tune (a) (b) (c) three CNN respectively.

Experimental Results:

The experimental results given by the authors are on average better than facetracker,panda-w, Panda-l good 8%,10%,3%. However, in Pre-train (a) (b), CNN used an additional imagenet dataset, and an additional celebfaces dataset was used for CNN in Pre-train (c). The author mentions panda-l useful to ground truth bounding box and landmark positions, but because I don't know facetracker,panda-w, Panda-l whether the three methods of comparison are useful to other datasets or other information, so it is not fair to give an opinion as to whether the results of the experiment are fairly good.

Results Analysis:

There are some interesting analysis in the article. The first is the analysis of the role of each of the neuron in the last CNN fully connected layer in (c). If the face images to each of the neuron response is divided into strong, medium, weak three categories, each class calculated an average faces, and finally found in fact every neuron have a semantic concept. As shown in the following:

The second analysis is (c) of CNN's last fully connected layer of the weight matrix, and each column vector (which I think is a row vector) is equivalent to a hyperplane that corresponds to a positive and negative sample that separates a attribute, Then all the column vectors are the hyperplanes that separate all attributes. These vectors are clustered with K-means, and finally, similar attributes are clustered together, suggesting that similar attributes hyperplane are similar.

Deep Learning paper Notes (3): Deep learning face Attributes in the Wild