目标：进行实时人脸识别，比https://github.com/ageitgey/face_recognition 中facerec_from_webcam_faster.py 的效果要稳、识别率更高


Our solution

•Face detection in video, through LSTM on video face dataset.

•Infrequent face recognition based on a small number of frames.


原链接solutions (blog)

https://medium.com/@ageitgey/machine-learning-is-fun-part-4-modern-face-recognition-with-deep-learning-c3cffc121d78

•Face detection on individual frames using HOG

•Frequent face recognition on individual frames


比较标准Comparison metric

•Better design —

•Speed (infrequent vs. each frame; network inference vs. HOG)

•Come up with some quantitative numbers.

•Consistency (inherent)

•Accuracy? Think about it.


image.png


实现思路：

1、detection用LSTM（上图右部分）

用video face detection dataset训一个lstm的network去detect视频中的人脸，代替原来的step 1（https://medium.com/@ageitgey/machine-learning-is-fun-part-4-modern-face-recognition-with-deep-learning-c3cffc121d78）hog

2、 step2 保留，还是进行posing and projecting face

3、上图左半部分， 每个方框代表一帧，之前所说的给一个人的所有bounding box一个id的意思是 我们使用前几帧去进行recognition，video里的人尽量保持不动 通过这几帧确定这个人后，以后的帧都可以锁定这个人。意思是不是每帧都做recognition，并且这些box相当于一个history， 假设这个人在20秒后出画面再进来 不用进行recognition还能认出这个人


导师给的：

Video face detection

•Data: multiple-faces detection

•Model: LSTM (each frame produce an output)

•We have done plenty of work on similar setting


Video face recognition

•Face detection in each frame: from the LSTM model.

•Transfer learning: different datasets (video face data & real time)

•Recognition would be based on a small number of frames.

•Infrequent: know box IDs.


综上：第一点需要做的就是找detection这块video的dataset，老师预计有这样训好的lstm 改一改也能用，他上面第三句话的意思是 我们实验室有这样做好的model是语音识别的 我改改也可以用。第二点是detection的数据recognition时能否用。原blog里后面的两个step他没有提，你可以看看有没有必要用。