The official Faster R-CNN code (written in MATLAB) is available here. If your goal is to reproduce the results in our NIPS 2015 paper, please use the official code.
This repository contains a Python reimplementation of the MATLAB code. This Python implementation is built on a fork of Fast R-CNN. There are slight differences between the two implementations. In particular, this Python port
- is ~10% slower at test-time, because some operations execute on the CPU in Python layers (e.g., 220ms / image vs. 200ms / image for VGG16)
- gives similar, but not exactly the same, mAP as the MATLAB version
- is not compatible with models trained using the MATLAB code due to the minor implementation differences
- includes approximate joint training that is 1.5x faster than alternating optimization (for VGG16) -- see these slides for more information
This repo is a fork of rbgirshick/py-faster-rcnn
which allows using Py-Faster-RCNN on Windows. Its branch windows
was forked from rbgirshick/py-faster-rcnn
's master branch on 20 May 2016, and contains minimal modifications that allow running training, evaluation, and deployment on Windows.
This README file was changed accordingly to reflect the changes in the installation and usage under Windows.
Note about Caffe: In this repo I use a different version of Caffe than rbgirshick/caffe-fast-rcnn
. I use kukuruza/caffe-fast-rcnn
, which is essentially the faster-rcnn
branch of rbgirshick/caffe-fast-rcnn
merged into the master branch of MSRCCS/caffe
repo. The merge was made in May 20, 2016. Therefore, an advantage of this repo over the original rbgirshick/py-faster-rcnn
is that a newer version of Caffe is used. See the installation section for details.
The work has been done while I was was a Microsoft Research intern in the Cloud Computing and Storage group.
By Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun (Microsoft Research)
The original Python implementation contains contributions from Sean Bell (Cornell) written during an MSR internship.
Please see the official README.md for more details.
Faster R-CNN was initially described in an arXiv tech report and was subsequently published in NIPS 2015.
Faster R-CNN is released under the MIT License (refer to the LICENSE file for details).
If you find Faster R-CNN useful in your research, please consider citing:
@inproceedings{renNIPS15fasterrcnn,
Author = {Shaoqing Ren and Kaiming He and Ross Girshick and Jian Sun},
Title = {Faster {R-CNN}: Towards Real-Time Object Detection
with Region Proposal Networks},
Booktitle = {Advances in Neural Information Processing Systems ({NIPS})},
Year = {2015}
}
- Requirements: software
- Requirements: hardware
- Basic installation
- Demo
- Beyond the demo: training and testing
- Usage
- Requirements for
Caffe
andpycaffe
(see: Caffe installation instructions) - Python 2.7+
- Python packages you might not have:
cython
,python-opencv
,easydict
- [Optional] MATLAB is required for official PASCAL VOC evaluation only. The code now includes unofficial Python evaluation code.
- For training smaller networks (ZF, VGG_CNN_M_1024) a good GPU (e.g., Titan, K20, K40, ...) with at least 3G of memory suffices
- For training Fast R-CNN with VGG16, you'll need a K40 (~11G of memory)
- For training the end-to-end version of Faster R-CNN with VGG16, 3G of GPU memory is sufficient (using CUDNN)
Type commands into Powershell.
-
Clone the Faster R-CNN repository, and checkout the required branch
git clone https://github.com/kukuruza/py-faster-rcnn chdir py-faster-rcnn git checkout win
We'll call the directory that you cloned Faster R-CNN into
FRCN_ROOT
. Note that Caffe is not a submodule any longer. -
Clone
caffe-fast-rcnn
repo insideFRCN_ROOT
and checkout the required branch. (Alternatively you can clonecaffe-fast-rcnn
into a different location and create a symbolic link.)chdir $FRCN_ROOT\.. git clone https://github.com/kukuruza/caffe mv caffe caffe-fast-rcnn chdir caffe-fast-rcnn git checkout win-faster-rcnn
-
Build the Cython modules
chdir $FRCN_ROOT\lib python setup.py build_ext --inplace cmd /c rmdir build /s /q
-
Build Caffe and pycaffe
Follow the instructions from https://github.com/Microsoft/caffe.
Make sure to compile WITH_PYTHON_LAYER
and build pycaffe
.
It is also recommended that you use CUDNN.
-
Download pre-computed Faster R-CNN detectors
chdir $FRCN_ROOT python data\scripts\fetch_faster_rcnn_models.py
This will populate the
$FRCN_ROOT\data
folder withfaster_rcnn_models
. Seedata\README.md
for details. These models were trained on VOC 2007 trainval.
After successfully completing basic installation, you'll be ready to run the demo.
To run the demo
chdir $FRCN_ROOT
python tools\demo.py
The demo performs detection using a VGG16 network trained for detection on PASCAL VOC 2007.
-
Download the training, validation, test data and VOCdevkit
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar
-
Extract all of these tars into one directory named
VOCdevkit
-
It should have this basic structure
$VOCdevkit\ # development kit $VOCdevkit\VOCcode\ # VOC utility code $VOCdevkit\VOC2007\ # image sets, annotations, etc. # ... and several other directories ...
-
Create symlinks for the PASCAL VOC dataset (need to run PowerShell as administrator)
chdir $FRCN_ROOT\data cmd /c mklink -d $VOCdevkit VOCdevkit2007
Using symlinks is a good idea because you will likely want to share the same PASCAL dataset installation between multiple projects.
-
[Optional] follow similar steps to get PASCAL VOC 2010 and 2012
-
[Optional] If you want to use COCO, please see some notes under
data\README.md
-
Follow the next sections to download pre-trained ImageNet models
Pre-trained ImageNet models can be downloaded for the three networks described in the paper: ZF and VGG16.
chdir $FRCN_ROOT
python data\scripts\fetch_imagenet_models.py
VGG16 comes from the Caffe Model Zoo, but is provided here for your convenience. ZF was trained at MSRA.
To train and test a Faster R-CNN detector using the alternating optimization algorithm from our NIPS 2015 paper, use experiments\scripts\faster_rcnn_alt_opt.py
.
Output is written underneath $FRCN_ROOT\output
.
chdir $FRCN_ROOT
python experiments\scripts\faster_rcnn_alt_opt.py [--GPU gpu_id] [--NET net] [...]
# GPU is the GPU id you want to train on
# NET in {ZF, VGG_CNN_M_1024, VGG16} is the network arch to use
# last optional arguments allow you to specify fast_rcnn.config options, e.g.
# EXP_DIR seed_rng1701 RNG_SEED 1701
("alt opt" refers to the alternating optimization training algorithm described in the NIPS paper.)
To train and test a Faster R-CNN detector using the approximate joint training method, use experiments\scripts\faster_rcnn_end2end.py
.
Output is written underneath $FRCN_ROOT\output
.
chdir $FRCN_ROOT
python experiments\scripts\faster_rcnn_end2end.py [--GPU gpu_id] [--NET net] [...]
# GPU is the GPU id you want to train on
# NET in {ZF, VGG_CNN_M_1024, VGG16} is the network arch to use
# last optional arguments allow you to specify fast_rcnn.config options, e.g.
# EXP_DIR seed_rng1701 RNG_SEED 1701
This method trains the RPN module jointly with the Fast R-CNN network, rather than alternating between training the two. It results in faster (~ 1.5x speedup) training times and similar detection accuracy. See these slides for more details.
Artifacts generated by the scripts in tools
are written in this directory.
Trained Fast R-CNN networks are saved under:
output\<experiment directory>\<dataset name>\
Test outputs are saved under:
output\<experiment directory>\<dataset name>\<network snapshot name>\