Dynamic Head

By Lin Song, Yanwei Li, Zhengkai Jiang, Zeming Li, Hongbin Sun, Jian Sun, Nanning Zheng.

This repo is an official implementation for "Fine-Grained Dynamic Head for Object Detection" (NeurIPS2020) on PyTorch framework.

Installation

Requirements

• Python >= 3.6
• PyTorch >= 1.6 and torchvision
• OpenCV, needed by demo and visualization
• pycocotools:
  • pip install cython
  • pip install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
• prodict:
  • pip install prodict
• GCC >= 4.9

Build from source

• git clone https://github.com/StevenGrove/DynamicHead
• cd DynamicHead
• python setup.py build develop

Prepare data

See datasets/README.md.

Usage

All the projects are placed in cvpods_playground. You need to select a project and enter the corresponding folder.

# For example
cd cvpods_playground/fcos.res50.1x

Training

# Running training procedure with specific GPU number
cvpods_train --gpu-nums <GPU_NUM> [optional arguments]

# Please refer to tools/train_net.py for more optional arguments

Inference

# Running inference procedure with specific GPU number
# The program will load the lastest model weights from ./log folder
cvpods_test --gpu-nums <GPU_NUM> [optional arguments]

# Running inference procedure with specific GPU number and model path
cvpods_test --gpu-nums <GPU_NUM> MODEL.WEIGHTS <model_path> [optional arguments]

# Please refer to tools/test_net.py for more optional arguments

Visualize your image [Optional]

# Please refer to demo/demo.py for more optional arguments
python ../../demo/demo.py --config <config_path> --input <input_path> --output <output_path> MODEL.WEIGHTS <model_path> [optional arguments]

Performance

Object Detection on COCO val set

All the results are based on ResNet-50 backbone. The FLOPs are calculated on the head only. "Low", "Mid" and "High" correspond to three configurations of different computational complex.

More Results is Coming

To do

• [ ] Easier installation based on pip
• [ ] Fast implementation for masked convolution
• [ ] Application on two-stage object detectors

Citation

Please cite the paper in your publications if it helps your research.

@inproceedings{song2020fine-grained,
    title = {Fine-Grained Dynamic Head for Object Detection},
    author = {Song, Lin and Li, Yanwei and Jiang, Zhengkai and Li, Zeming and Sun, Hongbin and Sun, Jian and Zheng, Nanning},
    booktitle = {Advances in Neural Information Processing Systems},
    year = {2020}
}

Please cite this project in your publications if it helps your research.

@misc{dynamichead,
    author = {Song, Lin},
    title = {DynamicHead},
    howpublished = {\url{https://github.com/StevenGrove/DynamicHead}},
    year ={2020}
}