dvlab-research / PanopticFCN

Fully Convolutional Networks for Panoptic Segmentation (CVPR2021 Oral)

PanopticFCN

Fully Convolutional Networks for Panoptic Segmentation

Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Liwei Wang, Zeming Li, Jian Sun, Jiaya Jia

[arXiv] [BibTeX]

This project provides an implementation for the CVPR 2021 Oral paper "Fully Convolutional Networks for Panoptic Segmentation" based on Detectron2. Panoptic FCN is a conceptually simple, strong, and efficient framework for panoptic segmentation, which represents and predicts foreground things and background stuff in a unified fully convolutional pipeline.

Installation

This project is based on Detectron2, which can be constructed as follows.

• Install Detectron2 following the instructions.
• Setup the dataset following the structure.
• Copy this project to /path/to/detectron2/projects/PanopticFCN

Training

To train a model with 8 GPUs, run:

cd /path/to/detectron2
python3 projects/PanopticFCN/train.py --config-file <config.yaml> --num-gpus 8

For example, to launch PanopticFCN training (1x schedule) with ResNet-50 backbone on 8 GPUs, one should execute:

cd /path/to/detectron2
python3 projects/PanopticFCN/train.py --config-file projects/PanopticFCN/configs/PanopticFCN-R50-1x.yaml --num-gpus 8

Evaluation

To evaluate a pre-trained model with 8 GPUs, run:

cd /path/to/detectron2
python3 projects/PanopticFCN/train.py --config-file <config.yaml> --num-gpus 8 --eval-only MODEL.WEIGHTS /path/to/model_checkpoint

Results

We provide the results on COCO val set with pretrained models. FPS is measured on a single V100 GPU.

Method	Backbone	Sched	PQ	SQ	RQ	AP	mIoU	FPS	download
PanopticFCN	R50	1x	41.1	79.8	49.9	32.2	41.5	12.4	model | metrics
PanopticFCN-400	R50	3x	41.0	81.0	49.6	30.7	43.6	22.5	model | metrics
PanopticFCN-512	R50	3x	42.3	81.1	51.2	32.4	43.2	19.8	model | metrics
PanopticFCN-600	R50	3x	42.7	81.2	51.5	33.6	43.9	17.5	model | metrics
PanopticFCN	R50	3x	43.6	81.4	52.5	34.4	43.6	12.8	model | metrics
PanopticFCN*	R50	3x	44.2	81.7	52.9	35.6	43.9	9.3	model | metrics

A faster version is also provided with higher threshold but similar PQ results, which shares the same model with the corresponding normal one. This version could be suitable for you if the final panoptic results are taken into consideration only.

Method	Backbone	Sched	PQ	SQ	RQ	AP	mIoU	FPS	download
PanopticFCN	R50	1x	41.1	79.8	49.9	30.2	41.4	13.6	model | metrics
PanopticFCN-400	R50	3x	40.8	81.1	49.4	28.9	43.5	26.1	model | metrics
PanopticFCN-512	R50	3x	42.3	81.1	51.2	30.7	43.2	22.0	model | metrics
PanopticFCN-600	R50	3x	42.7	80.8	51.4	31.6	43.9	19.1	model | metrics
PanopticFCN	R50	3x	43.6	81.4	52.5	32.4	43.6	13.5	model | metrics
PanopticFCN*	R50	3x	44.2	81.7	52.9	33.4	43.9	9.7	model | metrics

Citing PanopticFCN

Consider cite PanopticFCN in your publications if it helps your research.

@inproceedings{li2021panopticfcn,
  title={Fully Convolutional Networks for Panoptic Segmentation},
  author={Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia},
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2021}
}

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

@misc{PanopticFCN,
    author = {Yanwei Li},
    title = {PanopticFCN},
    howpublished = {\url{https://github.com/yanwei-li/PanopticFCN}},
    year ={2021}
}