GCNet for Object Detection

By Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, Han Hu.

This repo is a official implementation of "GCNet: Non-local Networks Meet Squeeze-Excitation Networks and Beyond" on COCO object detection based on open-mmlab's mmdetection. The core operator GC block could be find here. Many thanks to mmdetection for their simple and clean framework.

Update on 2020/12/07

The extension of GCNet got accepted by TPAMI (PDF).

Update on 2019/10/28

GCNet won the Best Paper Award at ICCV 2019 Neural Architects Workshop!

Update on 2019/07/01

The code is refactored. More results are provided and all configs could be found in configs/gcnet.

Notes: Both PyTorch official SyncBN and Apex SyncBN have some stability issues. During training, mAP may drops to zero and back to normal during last few epochs.

Update on 2019/06/03

GCNet is supported by the official mmdetection repo here. Thanks again for open-mmlab's work on open source projects.

Introduction

GCNet is initially described in arxiv. Via absorbing advantages of Non-Local Networks (NLNet) and Squeeze-Excitation Networks (SENet), GCNet provides a simple, fast and effective approach for global context modeling, which generally outperforms both NLNet and SENet on major benchmarks for various recognition tasks.

Citing GCNet

@article{cao2019GCNet,
  title={GCNet: Non-local Networks Meet Squeeze-Excitation Networks and Beyond},
  author={Cao, Yue and Xu, Jiarui and Lin, Stephen and Wei, Fangyun and Hu, Han},
  journal={arXiv preprint arXiv:1904.11492},
  year={2019}
}

Main Results

Results on R50-FPN with backbone (fixBN)

Results on R50-FPN with backbone (syncBN)

Results on stronger backbones

Notes

• GC denotes Global Context (GC) block is inserted after 1x1 conv of backbone.
• DCN denotes replace 3x3 conv with 3x3 Deformable Convolution in c3-c5 stages of backbone.
• r4 and r16 denote ratio 4 and ratio 16 in GC block respectively.
• Some of models are trained on 4 GPUs with 4 images on each GPU.

Requirements

• Linux(tested on Ubuntu 16.04)
• Python 3.6+
• PyTorch 1.1.0
• Cython
• apex (Sync BN)

Install

a. Install PyTorch 1.1 and torchvision following the official instructions.

b. Install latest apex with CUDA and C++ extensions following this instructions. The Sync BN implemented by apex is required.

c. Clone the GCNet repository.

 git clone https://github.com/xvjiarui/GCNet.git 

d. Compile cuda extensions.

cd GCNet
pip install cython  # or "conda install cython" if you prefer conda
./compile.sh  # or "PYTHON=python3 ./compile.sh" if you use system python3 without virtual environments

e. Install GCNet version mmdetection (other dependencies will be installed automatically).

python(3) setup.py install  # add --user if you want to install it locally
# or "pip install ."

Note: You need to run the last step each time you pull updates from github. Or you can run python(3) setup.py develop or pip install -e . to install mmdetection if you want to make modifications to it frequently.

Please refer to mmdetection install instruction for more details.

Environment

Hardware

• 8 NVIDIA Tesla V100 GPUs
• Intel Xeon 4114 CPU @ 2.20GHz

Software environment

• Python 3.6.7
• PyTorch 1.1.0
• CUDA 9.0
• CUDNN 7.0
• NCCL 2.3.5

Usage

Train

As in original mmdetection, distributed training is recommended for either single machine or multiple machines.

./tools/dist_train.sh <CONFIG_FILE> <GPU_NUM> [optional arguments]

Supported arguments are:

• --validate: perform evaluation every k (default=1) epochs during the training.
• --work_dir : if specified, the path in config file will be replaced.

Evaluation

To evaluate trained models, output file is required.

python tools/test.py <CONFIG_FILE> <MODEL_PATH> [optional arguments]

Supported arguments are:

• --gpus: number of GPU used for evaluation
• --out: output file name, usually ends wiht .pkl
• --eval: type of evaluation need, for mask-rcnn, bbox segm would evaluate both bounding box and mask AP.