Facial landmark detection based on convolution neural network.
This project is forked from here but implemented with caffe.
Here are some sample gifs extracted from video file showing the detection result compared with Dlib. The result of CNN is on the right side.
- Download 300-w dataset from here.
- Download 300-vw dataset from here or baidu cloud.
Extract video frame from 300vw:
cd dataset/300VW_Dataset_2015-12-14
bash ../extract_video_frame.sh
Randomly select 10% of extracted frame as train set as there are plenty of images looks the same. For example, two adjacent frames.
cd utils
python random_copy.py
Extract facial area from image:
python extract_face_from_ibug_image.py
Generate test and train list:
python file_list_generator.py
Generate HDF5 data:
python generate_hdf5.py
Start train:
cd model
./train.sh
This repo is a part of my deep learning series posts. For all the posts please refer to the following links.
为什么我决定采用深度学习实现面部特征点检测。阅读全文
解决问题所需的数据来源与对应的方法。阅读全文
从互联网获取的数据大多数情况下不是开箱即用的,这意味着我们需要对数据进行初步的整理,例如统计数据量、去除不需要的文件、必要的格式转换等。阅读全文
如何使用Python从22万张图片中提取检测人脸特征点的可用样本。阅读全文
将面部区域的图片与特征点位置一起打包成TensorFlow可用的TFRecord文件。阅读全文
如何使用TensorFlow构建一个属于你自己的神经网络模型。阅读全文
使用Estimator API时,导出适用于推演的网络模型的正确方法。阅读全文
如何通过CoreML在iPhone应用中使用TensorFlow模型。阅读全文