The codes are with the ICCV2015 paper "Convolutional Channel Features".

The codes include training and testing of a pedestrian detector on Caltech Pedestrian Dataset. We also provide our trained model for reproduction of the results on Caltech Pedestrian Detection Benchmark (reasonable, detailed) reported in the paper.

The codes are written in MATLAB, dependent on Caffe and Piotr's Computer Vision Matlab Toolbox. Codes are tested on Linux 12.04.3 LTS with 128GB memory and a Titan Z GPU.

• Make the provided Caffe version with matCaffe interface
• Download VGG-16 CaffeModel to ./data/CaffeNets/
• Download Caltech Pedestrian Dataset and set it up properly with codes in ./data/code3.2.1

• Run ./runDetect.m, and detection results will be saved as allBBs.mat

• Run modified ./toolbox-master/detector/acfDemoCal.m to train an ACF detector and save the collected samples for training CCF detector
• Run ./getFeat_train.m to extract features of all samples
• Run ./trainModel.m to train the boosting forest model
• Run ./runDetect.m using your trained model
• (Evaluation) Please refer to ./toolbox-master/detector/acfTest.m and ./data/code3.2.1/dbEval.m

• Power law can accelerate the detection time by a large margin with unnoticeable performance decrease. To our knowledge, it holds for CCF on AFW face dataset but doesn't hold on Caltech pedestrian dataset. (see the paper for more details)
• You can check whether the power law holds for specific feature type on specific dataset by running ./power_law/getMean.m and ./power_law/getLambda.m consecutively.

If you use our codes or model in your research, we are grateful if you cite the paper:

@inproceedings{binyang15ccf,
  Author    = {Bin Yang and
               Junjie Yan and
               Zhen Lei and
               Stan Z. Li},
  Title     = {Convolutional Channel Features},
  Booktitle = {Proceedings of the IEEE International Conference
               on Computer Vision (ICCV)},
  Year      = {2015}
}

Great gratitude is presented to

• Piotr Dollar's toolbox
• Caffe team
• VGG team
• NVIDIA Corporation