CenterDisks

Repository for the paper
by Katia Jodogne-del Litto¹, Guillaume-Alexandre Bilodeau¹.
¹ Polytechnique Montréal

Abstract

CenterDisks: Real-time instance segmentation with disk covering

Example results

Initial image	Disks covering

Requirements:

• python 3.8
• pytorch (1.8.0, cuda 10.2)
• mmdetection (for DCN)
• various common packages (opencv, numpy...)

Folder organization:

• experiments/: scripts for the experiments.
• Data path expected usually is /store/datasets/DATASET_NAME (changeable in code)
• src/lib/trains/gaussiandet.py is the training file
• src/lib/datasets/sample/gaussiandet.py is the sampling file
• src/lib/detectors/gaussiandet.py is the detector file

Help

For general debugging and help to run the scripts:

• This code is built upon: https://github.com/xingyizhou/CenterNet and https://github.com/hu64/CenterPoly
• The pre-trained weights are all available at this location

Results:

Datasets (test sets)	AP	AP50%	Runtime (s)	weights
cityscapes	7.36	25.89	0.040	link
KITTI	11.75	37.24	0.032	link
IDD	6.70	21.40	0.033	link

Models

https://polymtlca0-my.sharepoint.com/:f:/g/personal/katia_jodogne--del-litto_polymtl_ca/ErwvXbsztOpNvtqLw-bWZakBSZV4XtgJyeEXXr-Z2nARQw?e=3uDYa0

Acknowledgements

The code for this paper is mainly built upon CenterPoly and CenterNet, we would therefore like to thank the authors for providing the source code of their paper. We also acknowledge the support of the Natural Sciences and Engineering Research Council of Canada (NSERC), and the support of IVADO [MSc-2022-4713306544].

License

CenterDisks is released under the MIT License. Portions of the code are borrowed from CenterPoly, CenterNet, CornerNet (hourglassnet, loss functions), dla (DLA network), mmdetection(deformable convolutions), and cityscapesScripts (cityscapes dataset evaluation). Please refer to the original License of these projects (See NOTICE).