DNLNet for Object Detection

By Minghao Yin, Zhuliang Yao, Yue Cao, Xiu Li, Zheng Zhang, Stephen Lin, Han Hu.

This repo is a official implementation of "Disentangled Non-Local Neural Networks" on COCO object detection based on open-mmlab's mmdetection. Many thanks to mmdetection for their simple and clean framework.

Introduction

DNLNet is initially described in arxiv. It is still in progress.

Citing DNLNet

@misc{yin2020disentangled,
    title={Disentangled Non-Local Neural Networks},
    author={Minghao Yin and Zhuliang Yao and Yue Cao and Xiu Li and Zheng Zhang and Stephen Lin and Han Hu},
    year={2020},
    booktitle={ECCV}
}

Main Results

Results on R50-FPN with backbone (syncBN)

Results on stronger backbones

• On going

Notes

• NL denotes Non-local block block is inserted after 1x1 conv of backbone.
• GC denotes Global Context (GC) block is inserted after 1x1 conv of backbone.
• SNL denotes Simplified Non-local block block is inserted after 1x1 conv of backbone.
• DNL denotes Disentangled Non-local block block is inserted after 1x1 conv of backbone.
• r4 denotes ratio 4 in GC block.
• c4 and c5_all denote insert context block at stage c4's last residual block and c5's all blocks, respectively.
• Most models are trained on 16 GPUs with 4 images on each GPU.

Requirements

• Linux(tested on Ubuntu 16.04)
• Python 3.6+
• Cython
• PyTorch 1.1.0
• CUDA 9.0
• CUDNN 7.0
• NCCL 2.3.5
• apex
• inplace_abn

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 inplace_abn implemented by apex is required.

c. Clone the DNLNet repository.

 git clone https://github.com/Howal/DNL-Object-Detection.git

d. Install DNLNet version mmdetection (other dependencies will be installed automatically).

cd DNL-Object-Detection
python(3) setup.py build develop  # add --user if you want to install it locally
# or "pip install -e -v ."

Please refer to mmdetection install instruction for more details.

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 <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.