Home

Awesome

Training Region-based Object Detectors with Online Hard Example Mining

By Abhinav Shrivastava, Abhinav Gupta, Ross Girshick

Introduction

Online Hard Example Mining (OHEM) is an online bootstrapping algorithm for training region-based ConvNet object detectors like Fast R-CNN. OHEM

OHEM was initially presented at CVPR 2016 as an Oral Presentation. For more details, see the arXiv tech report.

License

This code is released under the MIT License (refer to the LICENSE file for details).

Citing

If you find this useful in your research, please consider citing:

@inproceedings{shrivastavaCVPR16ohem,
    Author = {Abhinav Shrivastava and Abhinav Gupta and Ross Girshick},
    Title = {Training Region-based Object Detectors with Online Hard Example Mining},
    Booktitle = {Conference on Computer Vision and Pattern Recognition ({CVPR})},
    Year = {2016}
}

Disclaimer

This implementation is built on a fork of Faster R-CNN Python code (here), which in turn builds on Fast R-CNN (here). Please cite the appropriate papers depending on which part of the code and/or model you are using.

Results

training datatest datamAP (paper)mAP (this repo)
Fast R-CNN (FRCN)VOC 07 trainvalVOC 07 test66.967.6
FRCN with OHEMVOC 07 trainvalVOC 07 test69.971.5
FRCN, +M, +BVOC 07 trainvalVOC 07 test72.4
FRCN with OHEM, +M, +BVOC 07 trainvalVOC 07 test75.1
FRCNVOC 07 trainval + 12 trainvalVOC 07 test70.0
FRCN with OHEMVOC 07 trainval + 12 trainvalVOC 07 test74.675.5
FRCN with OHEM, +M, +BVOC 07 trainval + 12 trainvalVOC 07 test78.9
FRCNVOC 12 trainvalVOC 12 test65.7
FRCN with OHEMVOC 12 trainvalVOC 12 test69.8
FRCN with OHEM, +M, +BVOC 12 trainvalVOC 12 test72.9
FRCNVOC 07 trainval&test + 12 trainvalVOC 12 test68.4
FRCN with OHEMVOC 07 trainval&test + 12 trainvalVOC 12 test71.9
FRCN with OHEM, +M, +BVOC 07 trainval&test + 12 trainvalVOC 12 test76.3
FRCN with OHEM, +M, +Babove + COCO 14 trainvalVOC 12 test80.1

Note: All methods above use the VGG16 network. mAP (paper) is the mAP reported in the paper. mAP (this repo) is the mAP reproduced by this codebase.

Legend: +M: using multi-scale for training and testing, +B: multi-stage bounding box regression. See the paper for details.

Released

Sometime in the future

Contents

  1. Requirements: software
  2. Requirements: hardware
  3. Basic installation
  4. Demo
  5. Beyond the demo: training and testing
  6. Usage
  7. FAQ regarding Faster R-CNN support

Requirements: software

  1. Requirements for Caffe and pycaffe (see: Caffe installation instructions)

Note: Caffe must be built with support for Python layers!

# In your Makefile.config, make sure to have this line uncommented
WITH_PYTHON_LAYER := 1
# Unrelatedly, it's also recommended that you use CUDNN
USE_CUDNN := 1

You can download Ross's Makefile.config for reference. 2. Python packages you might not have: cython, python-opencv, easydict, `yaaml' 3. [Optional] MATLAB is required for official PASCAL VOC evaluation only. The code now includes unofficial Python evaluation code.

Requirements: hardware

  1. For training smaller networks (VGG_CNN_M_1024) a good GPU (e.g., Titan, K20, K40, ...) with at least 4G of memory suffices
  2. For training VGG16, you'll need a K40 or Titan X (or better).

Installation (similar to Fast(er) R-CNN)

  1. Clone the OHEM repository
# Make sure to clone with --recursive
git clone --recursive https://github.com/abhi2610/ohem.git
  1. We'll call the directory that you cloned OHEM into OHEM_ROOT

    Ignore notes 1 and 2 if you followed step 1 above.

    Note 1: If you didn't clone OHEM with the --recursive flag, then you'll need to manually clone the caffe-fast-rcnn submodule:

    git submodule update --init --recursive
    

    Note 2: The caffe-fast-rcnn submodule needs to be on the faster-rcnn branch (or equivalent detached state). This will happen automatically if you followed step 1 instructions.

  2. Build the Cython modules

    cd $OHEM_ROOT/lib
    make
    
  3. Build Caffe and pycaffe

    cd $OHEM_ROOT/caffe-fast-rcnn
    # 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 pycaffe
    
  4. Download pre-computed Fast R-CNN detector trained with OHEM using VGG16 and VGG_CNN_M_1024 networks.

    cd $OHEM_ROOT
    ./data/scripts/fetch_fast_rcnn_ohem_models.sh
    

    This will populate the $OHEM_ROOT/data folder with a fast_rcnn_ohem_models folder which contains VGG16 and VGG_CNN_M_1024 models (Fast R-CNN detectors trained with OHEM). The format will be fast_rcnn_ohem_models/TRAINING_SET/MODEL_FILE.

These models were re-trained using this codebase and achieve slightly better performance (see this Table). In particular, on the standard split, VGG_CNN_M_1024 model gets 62.8 mAP (compared to 62.0 mAP reported in paper) and VGG16 model gets 71.5 mAP (compared to 69.9 mAP). All models from the paper will be released soon.

Installation for training and testing models

  1. 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
  1. Extract all of these tars into one directory named VOCdevkit
tar xvf VOCtrainval_06-Nov-2007.tar
tar xvf VOCtest_06-Nov-2007.tar
tar xvf VOCdevkit_08-Jun-2007.tar
  1. It should have this basic structure
  $VOCdevkit/                           # development kit
  $VOCdevkit/VOCcode/                   # VOC utility code
  $VOCdevkit/VOC2007                    # image sets, annotations, etc.
  # ... and several other directories ...
  ```

4. Create symlinks for the PASCAL VOC dataset

```Shell
  cd $OHEM_ROOT/data
  ln -s $VOCdevkit VOCdevkit2007
  ```
  Using symlinks is a good idea because you will likely want to share the same PASCAL dataset installation between multiple projects.
5. [Optional] follow similar steps to get PASCAL VOC 2010 and 2012
6. Follow the next sections to download pre-trained ImageNet models

*COCO instructions and models will be released soon.*

### Download pre-trained ImageNet models

Pre-trained ImageNet models can be downloaded for the two networks described in the paper: VGG_CNN_M_1024 and VGG16.

```Shell
cd $OHEM_ROOT
./data/scripts/fetch_imagenet_models.sh

Models come from the Caffe Model Zoo, but are provided here for your convenience..

Usage

To train a Fast R-CNN detector using the OHEM algorithm on voc_2007_trainval, use experiments/scripts/fast_rcnn_ohem.sh. See experiments/scripts/ directory for other scripts. Output is written underneath $OHEM_ROOT/output.

cd $OHEM_ROOT
./experiments/scripts/fast_rcnn_ohem.sh [GPU_ID] [NET] [--set ...]
# GPU_ID is the GPU you want to train on
# NET in {VGG16, VGG_CNN_M_1024} is the network arch to use
# --set ... allows you to specify fast_rcnn.config options, e.g.
#   --set EXP_DIR seed_rng1701 RNG_SEED 1701

Artifacts generated by the scripts in tools are written in this directory.

Trained Fast R-CNN networks with OHEM are saved under:

output/<experiment directory>/<dataset name>/

Test outputs are saved under:

output/<experiment directory>/<dataset name>/<network snapshot name>/

The VGG_CNN_M_1024 model should get ~62.8 mAP and VGG16 model should get ~71.5 mAP. For reference, you can download my logs from here.

FAQ regarding Faster R-CNN support

I have received a lot of queries regarding using OHEM with Faster R-CNN. I have not spent too much time combining OHEM with Faster R-CNN yet. Some researchers have informed me that OHEM works well in the 'alternating optimization' setup, but not so much with the 'end to end learning' setup. I hope to try and release the support for Faster R-CNN in the coming months. If you would like an update when I release it, send me an email.

Also, the authors of R-FCN succesfully used OHEM with R-FCN and Faster R-CNN; you might find their codebase helpful.