Official PyTorch implementation of the CVPR 2020 Oral Real-Time Panoptic Segmentation from Dense Detections by the ML Team at Toyota Research Institute (TRI), cf. References below.
git clone https://github.com/TRI-ML/realtime_panoptic.git
cd realtime_panoptic
make docker-build
To verify your installation, you can also run our simple test run to conduct inference on 1 test image using our Cityscapes pretrained model:
make docker-run-test-sample
Now you can start a docker container with interactive mode:
make docker-start
We provide demo code to conduct inference on Cityscapes pretrained model.
python scripts/demo.py --config-file <config.yaml> --input <input_image_file> \
--pretrained-weight <checkpoint.pth>
Simple user example using our pretrained model previded in the Models section:
python scripts/demo.py --config-file ./configs/demo_config.yaml --input media/figs/test.png --pretrained-weight cvpr_realtime_pano_cityscapes_standalone_no_prefix.pth
| Model | PQ | PQ_th | PQ_st |
|---|---|---|---|
| ResNet-50 | 58.8 | 52.1 | 63.7 |
The source code is released under the MIT license.
Rui Hou*, Jie Li*, Arjun Bhargava, Allan Raventos, Vitor Guizilini, Chao Fang, Jerome Lynch, Adrien Gaidon, [paper], [oral presentation], [teaser]
@InProceedings{real-time-panoptic,
author = {Hou, Rui and Li, Jie and Bhargava, Arjun and Raventos, Allan and Guizilini, Vitor and Fang, Chao and Lynch, Jerome and Gaidon, Adrien},
title = {Real-Time Panoptic Segmentation From Dense Detections},
booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}