Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Neural Networks.
This project provide you a method to update multi-task-loss for multi-input source.
- Ubuntu 14.04 or 16.04
- caffe && pycaffe: https://github.com/BVLC/caffe
- cPickle && cv2 && numpy
The training data generate process can refer to Seanlinx/mtcnn
Sample almost similar to Seanlinx's can be found in
step1. Download Wider Face Training part only from Official Website and unzip to replace
gen_12net_data.pyto generate 12net training data. Besides,
gen_net_imdb.pyprovide you an example to build imdb, Remember changing and adding new params.
gen_12net_hard_example.pyto generate hard sample. Run
gen_24net_data.py. Combine these output and generate imdb.
step4. Similar to last step, Run
gen_24net_hard_example.pyto generate hard sample. Run
gen_48net_data.py. Combine these output and generate imdb.
Strongly suggest readers generate training data themselves. The sample training data of 12net and 24net is available( Too big for Baidu Drive to upload 48net) by sending Email
The main idea is block backward propagation for different task
The Q&A bellow can solve most of your problem.
Q1: What data base do you use?
A1: Similar to official paper, Wider Face for detection and CelebA for alignment.
Q2: What is "12(24/48)net-only-cls.caffemodel" file for?
A2: Provide a initial weigh to train. Since caffe's initial weigh is random, a bad initial weigh may take a long ran to converge even might overfit before that.
Q3: Why preprocess images by minus 128?
A3: Separating data from (0,+) to (-,+), can make converge faster and more accurate. Refer to Batch Normalization Accelerating Deep Network Training by Reducing Internal Covariate Shift
Q4: Do you implement OHEM(Online-Hard-Example-Mining)?
A4: No. OHEM is used when training data is not that much big. Refer to faster-rcnn's writer RBG's paper
Q5: Ratio positive/negative samples for 12net?
A5: This caffemodel used neg:pos=3:1. Because 12net's function is to eliminate negative answers, similar to exclusive method, we should learn more about negative elininate the wrong answer.
Q6: Why your stride is different to official?
A6: If you input a (X,X) image, the output Y = (X-11)/2. Every point on output represent a ROI on input. The ROI's left side moving range = (0, X-12) on input, and (0, Y-1) on output. So that stride = (X-12)/(Y-1) ≈≈ 2 in this net.
Q7: What is roi(cls/pts).imdb used for?
A7: Use imdb can feed training data into training net faster. Instead of random search data from the hard-disk, reading data from a large file once to memory will save you a lot of time.
imdb was created by python module-cPickle.
Q8: What is
tools_matrix.py different from
A8: Matrix version use linear matrix to make calculation faster(160ms on FDDB). If you are green hand in this area, read Non-Matrix version to understand each process.
Q9: I want your training data for emergency use. How to use them? How to train? ...
CongWeilin updated in 2017/3/5
tools_matrix.py to make calculate faster, about 160ms/image.