rbgirshick/DeepPyramid

DeepPyramid

DeepPyramid is a simple toolkit for building feature pyramids from deep convolutional networks. The DeepPyramid data structure is nearly identical to the HOG feature pyramid created by the featpyramid.m function in the voc-dpm code.

References

This code was used in our tech report about the relationship between deformable part models and convolutional networks.

@article{girshick14dpdpm,
    author    = {Ross Girshick and Forrest Iandola and Trevor Darrell and Jitendra Malik},
    title     = {Deformable Part Models are Convolutional Neural Networks},
    journal   = {CoRR},
    year      = {2014},
    volume    = {abs/1409.5403},
    url       = {http://arxiv.org/abs/1409.5403},
    year      = {2014}
}

Installation

0. Prerequisites
1. MATLAB (tested with 2014a on 64-bit Linux)
2. Caffe's prerequisites
3. Install Caffe (this is the most complicated part)
4. Follow the Caffe installation instructions
5. Let's call the place where you installed caffe $CAFFE_ROOT (you can run export CAFFE_ROOT=$(pwd))
6. Important: Make sure to compile the Caffe MATLAB wrapper, which is not built by default: make matcaffe
7. Important: Make sure to run cd $CAFFE_ROOT/data/ilsvrc12 && ./get_ilsvrc_aux.sh to download the ImageNet image mean
8. DeepPyramid has been tested with master and dev at the time of this writing
9. Get DeepPyramid
10. git clone https://github.com/rbgirshick/DeepPyramid.git
11. If you haven't installed R-CNN, you'll need to download its models
12. Copy R-CNN's non-finetuned ImageNet network <rcnnpath>/data/caffe_nets/ilsvrc_2012_train_iter_310k to <deeppyramidpath>/data/caffe_nets/ilsvrc_2012_train_iter_310k (or just create a symlink).

Usage

1. Run matlab from inside the DeepPyramid code directory
2. Add the matcaffe mex function to your path (addpath /path/to/caffe/matlab/caffe)
3. Run the demo demo_deep_pyramid

Uses

DeepPyramid can be used for implementing DPMs on deep convolutional network features, rather than HOG features. It can also be used whenever you need a dense multiscale pyramid of image features.

Caveats

The implementation is designed to be simple and as a result is very inefficient. There are a variety of ways to speed it up, and they will be done in the future. For now, it takes about 0.5 to 0.6 seconds to compute a feature pyramid on an NVIDIA Titan GPU, which is acceptable.