afcarl / recurrent-face-alignment Goto Github PK

This is a quick demo for:

"A Recurrent Encoder-Decoder Network for Sequential Face Alignment"

Xi Peng, Rogerio S. Feris, Xiaoyu Wang, Dimitris N. Metaxas

European Conference on Computer Vision (ECCV), Amsterdam, 2016. (Oral)

Download PDF, Slides, Poster

1. Clone/Download the project to recurrent-face-alignment/

2. Download folders model/ and data/ from Google Drive

3. Copy folders model/ and data/ into recurrent-face-alignment/

4. Edit TrackVideoDemo.py to set (a) path/to/caffe/python/, and (b) video names to be tracked

5. python TrackDemo.py

6. Check tracking results in recurrent-face-alignment/result/

4. Edit DetectImageDemo.py to set (a) path/to/caffe/python/, and (b) image folder to be detected

5. python DetectImageDemo.py

For research convenience, we split video into frames using ffmpeg.

The tracker need the bbox of the face at the first frame for initialization.

The detector need img_bbox.txt that each line has 5 tokens: path/to/image left top right bottom.

[left top right bottom] is the bbox of detected face.

caffe: any version that support batch normalization layer (such as SegNet). We will relase our distributed caffe version soon.

python 2.7.

@InProceedings{PengECCV16,
author = "Peng, Xi and Feris, Rogerio S.and Wang, Xiaoyu and Metaxas, Dimitris N.",
title = "A Recurrent Encoder-Decoder Network for Sequential Face Alignment",
booktitle = "European Conference on Computer Vision (ECCV)",
year = "2016",
pages="38--56"}