Deep Patch Learning for Weakly Supervised Object Classification and Discovery

By Peng Tang, Xinggang Wang, Zilong Huang, Xiang Bai, and Wenyu Liu.

Introduction

Deep Patch Learning (DPL) is a fast framework for object classification and discovery with deep ConvNets.

It achieves state-of-the-art performance on object classification (Pascal VOC 2007 and 2012), and very competitive results on object discovery.
Our code is written by C++ and Python, based on Caffe and fast r-cnn.

The paper has been accepted by Pattern Recognition. For more details, please refer to our paper (also available at arxiv).

If you are focusing on weakly supervised object detection (or object discovery), you can also see our recent CVPR2017 work OICR.

Architecture

DPL architecture

Results

	VOC2007 test mAP (classification)	VOC2007 trainval CorLoc (discovery)	VOC2012 test mAP (classification)	VOC2012 trainval CorLoc (discovery)
DPL-AlexNet	85.3	43.5	84.4	48.7
DPL-VGG16	92.7	45.4	92.5	51.0

Visualizations

Some pattern visualization results

Some pattern visualization results.

Some detection visualization results

Some detection visualization results.

License

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

Citing DPL

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

@article{tang2017deep,
    author = {Tang, Peng and Wang, Xinggang and Huang, Zilong and Bai, Xiang and Liu, Wenyu},
    title = {Deep Patch Learning for Weakly Supervised Object Classification and Discovery},
    journal = {Pattern Recognition},
    volume = {},
    pages = {},
    year = {2017}
}

Requirements: software
Requirements: hardware
Basic installation
Installation for training and testing
Extra Downloads (selective search)
Extra Downloads (ImageNet models)
Usage
Trained models

Requirements: software

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
```
Python packages you might not have: cython, python-opencv, easydict
MATLAB

Requirements: hardware

NVIDIA GTX TITANX (~12G of memory)

Installation (sufficient for the demo)

Clone the DPL repository

# Make sure to clone with --recursive
git clone --recursive https://github.com/ppengtang/dpl.git

Build the Cython modules
```
cd $DPL_ROOT/lib
make
```

Build Caffe and pycaffe

cd $DPL_ROOT/caffe-dpl
# 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 all -j 8 && make pycaffe

Installation for training and testing

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/voc2012/VOCdevkit_18-May-2011.tar

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_18-May-2011.tar

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
```
cd $DPL_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.
[Optional] follow similar steps to get PASCAL VOC 2012.
You should put the generated proposal data under the folder $DPL_ROOT/data/selective_search_data, with the name "voc_2007_trainval.mat", "voc_2007_test.mat", just as the form of fast-rcnn.
The pre-trained models are all available in the Caffe Model Zoo. You should put it under the folder $DPL_ROOT/data/imagenet_models, just as the form of fast-rcnn.

Download pre-computed Selective Search object proposals

Pre-computed selective search boxes can also be downloaded for VOC2007 and VOC2012.

cd $DPL_ROOT
./data/scripts/fetch_selective_search_data.sh

This will populate the $DPL_ROOT/data folder with selective_selective_data. (The script is copied from the fast-rcnn).

Download pre-trained ImageNet models

Pre-trained ImageNet models can be downloaded.

cd $DPL_ROOT
./data/scripts/fetch_imagenet_models.sh

These models are all available in the Caffe Model Zoo, but are provided here for your convenience. (The script is copied from the fast-rcnn).

Usage

Train a DPL network. For example, train a VGG16 network on VOC 2007 trainval:

./tools/train_net.py --gpu 1 --solver models/VGG16/solver.prototxt \
    --weights data/imagenet_models/$VGG16_model_name --iters 40000

Test a DPL network. For example, test the VGG 16 network on VOC 2007 test:

Classification

./tools/test_net.py --gpu 1 --def models/VGG16/test_cls.prototxt \
    --net output/default/voc_2007_trainval/vgg16_dpl_iter_40000.caffemodel

Discovery

./tools/test_net.py --gpu 1 --def models/VGG16/test_det.prototxt \
  --net output/default/voc_2007_trainval/vgg16_dpl_iter_40000.caffemodel \
  --imdb voc_2007_trainval --task det

Test output is written underneath $DPL_ROOT/output.

Evaluation

To get results, put the results under the folder $VOCdevkit/results/VOC2007/Main.

For classification, run the matlab code eval_classification.m

For discovery, run the matlab code eval_discovery.m

Our trained models

The models trained on PASCAL VOC 2007 can be downloaded from here.

And on PASCAL VOC 2012 can be downloaded from here.