Caffe and Python implementation of Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks.
Set up environment and copy C++ layer code to Caffe's source code tree.
$ export PYTHONPATH=/path/to/Joint-Face-Detection-and-Alignment:$PYTHONPATH
$ export CAFFE_HOME=/path/to/caffe
$ sh layers/copy.sh
Compile Caffe following its document.
Download dataset WIDER, CelebA and FDDB. Put them in data directory like below.
data
├── CelebA
│ └── img_celeba
├── fddb
│ ├── FDDB-folds
│ ├── images
│ │ ├── 2002
│ │ └── 2003
│ └── result
│ └── images
└── WIDER
├── wider_face_split
├── WIDER_test
├── WIDER_train
└── WIDER_val
I have write a matlab script to extract WIDER FACE info from matlab mat file to txt file.
Prepare data and train network follow the commands in train.sh.
Test the model with demo.py for simple detection and fddb.py for FDDB benchmark.
Since pNet may output many bboxes for rNet and Caffe's Blob never realloc the memory if your new data is smaller, this makes Blob only grow the memory and never reduce, which looks like a memory leak. It is fine for most cases but not for our case. You may modify src/caffe/blob.cpp if you encounter the memory issue.
template <typename Dtype>
void Blob<Dtype>::Reshape(const vector<int>& shape) {
/* some code */
if (count_ > capacity_) { // never reduce the memory here
capacity_ = count_;
data_.reset(new SyncedMemory(capacity_ * sizeof(Dtype)));
diff_.reset(new SyncedMemory(capacity_ * sizeof(Dtype)));
}
}template <typename Dtype>
void Blob<Dtype>::Reshape(const vector<int>& shape) {
/* some code */
if (count_ != capacity_) { // make a new data buffer
capacity_ = count_;
data_.reset(new SyncedMemory(capacity_ * sizeof(Dtype)));
diff_.reset(new SyncedMemory(capacity_ * sizeof(Dtype)));
}
}