DeepPyramid DPM

Introduction

The DeepPyramid DPM repository is code for our CVPR 2015 paper Deformable Part Models are Convolutional Neural Networks. This work was completed while Ross Girshick was a postdoc at UC Berkeley (prior to joining Microsoft Research). It implements Latent SVM training of deformable part models (DPMs) on top of deep feature pyramids.

@inproceedings{girshick2015dpdpm,
    Author    = {Ross Girshick and Forrest Iandola and
                 Trevor Darrell and Jitendra Malik},
    Title     = {Deformable Part Models are
                 Convolutional Neural Networks},
    Booktitle = {Proceedings of the IEEE Conference on
                 Computer Vision and Pattern Recognition ({CVPR})},
    Year      = {2015}
}

License

Deep Pyramid DPM is released under the MIT License (refer to the LICENSE file for details). Substantial parts of the code come from DPMv5, which is also under the MIT License (see COPYING.DPMv5 for details).

Requirements: software

Requirements for Caffe and matcaffe (see: Caffe installation instructions)

You can download my Makefile.config for reference. 2. MATLAB (tested on R2014a)

Requirements: hardware

A good GPU, such as a GTX Titan, K20, etc.

Installation

Clone the repository

# Make sure to clone with --recursive
git clone --recursive https://github.com/rbgirshick/dp-dpm.git

We'll call the directory that you cloned DeepPyramid DPM into DPDPM_ROOT

Build Caffe and matcaffe

cd $DPDPM_ROOT/caffe
# Now follow the Caffe installation instructions here:
#   http://caffe.berkeleyvision.org/installation.html

# If you're experienced with Caffe and have all of the requirements installed
# and your Makefile.config in place, then simply do:
make -j8 && make matcaffe
# Replace 8 with your favorite number of compile threads

Download the pre-computed ImageNet model
```
cd $DPDPM_ROOT
./data/scripts/fetch_imagenet_model.sh
```
This will populate the $DPDPM_ROOT/data/caffe_nets folder with CaffeNet.v2.caffemodel.

Symlink to the PASCAL VOC 2007 dataset

Note: for more information on installing VOC 2007, go here.

cd $DPDPM_ROOT
# Replace /path/to with the actual path to VOC 2007
ln -s /path/to/VOC2007/VOCdevkit cachedir/VOC2007/VOCdevkit

Build the MATLAB/mex code
```
# Start matlab
matlab
>> compile
```

Usage

To train and test a detector:

cd $DPDPM_ROOT
matlab
>> % run these commands inside matlab
>> precompute_feat_pyramids() % this will take several hours
>> pascal('bicycle', 3) % replace 'bicycle' with any PASCAL class

Executing precompute_feat_pyramids() is optional. If you don't call this function then features will be cached while training and testing for the first time (making training and testing much slower).

jwu76033/dp-dpm