Kev1n3zz/DGMN2

消息传递网络，可变的图卷积神经网络

DGMN2

This repository contains the implementation of Dynamic Graph Message Passing Networks.

Main results

Image Classification

ImageNet-1K

Method	Params (M)	FLOPs (G)	Top1 Acc (%)
DGMN2-Tiny	12.1	2.3	78.7
DGMN2-Small	21.0	4.3	81.7
DGMN2-Medium	35.8	7.1	82.5
DGMN2-Large	48.3	10.4	83.3

Object Detection

COCO validation set

Method	Backbone	Lr schd	box AP	mask AP
RetinaNet	DGMN2-Tiny	1x	39.7	-
RetinaNet	DGMN2-Small	1x	42.5	-
RetinaNet	DGMN2-Medium	1x	43.7	-
RetinaNet	DGMN2-Large	1x	44.7	-
Mask R-CNN	DGMN2-Tiny	1x	40.2	37.2
Mask R-CNN	DGMN2-Small	1x	43.4	39.7
Mask R-CNN	DGMN2-Medium	1x	44.4	40.2
Mask R-CNN	DGMN2-Large	1x	46.2	41.6
Deformable DETR	DGMN2-Tiny	50e	44.4	-
Deformable DETR	DGMN2-Small	50e	47.3	-
Deformable DETR	DGMN2-Medium	50e	48.4	-
Deformable DETR+	DGMN2-Small	50e	47.3	-
Sparse R-CNN	DGMN2-Small	3x	48.2	-

Semantic Segmentation

Cityscapes validation set

Method	Backbone	Iters	mIoU	mIoU (ms + flip)
Semantic FPN	DGMN2-Tiny	40K	78.09	79.40
Semantic FPN	DGMN2-Small	40K	80.65	81.58
Semantic FPN	DGMN2-Medium	40K	80.60	81.79
Semantic FPN	DGMN2-Large	40K	81.75	82.64
SETR-Naive	DGMN2-Tiny	40K	77.23	78.23
SETR-Naive	DGMN2-Small	40K	80.31	81.04
SETR-Naive	DGMN2-Medium	40K	80.83	81.39
SETR-Naive	DGMN2-Large	40K	81.80	82.61
SETR-PUP	DGMN2-Tiny	40K	78.25	79.26
SETR-PUP	DGMN2-Small	40K	79.78	80.73
SETR-PUP	DGMN2-Medium	40K	80.96	81.80
SETR-PUP	DGMN2-Large	40K	81.58	82.27
SETR-MLA	DGMN2-Tiny	40K	78.25	79.32
SETR-MLA	DGMN2-Small	40K	80.79	81.62
SETR-MLA	DGMN2-Medium	40K	81.09	82.00
SETR-MLA	DGMN2-Large	40K	81.55	81.98

Getting Started

• For image classification, please see classification.
• For object detection, please see detection.
• For semantic segmentation, please see segmentation.

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.

Reference

@inproceedings{zhang2020dynamic,
  title={Dynamic Graph Message Passing Networks},
  author={Zhang, Li and Xu, Dan and Arnab, Anurag and Torr, Philip H.S.},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2020}
}

@article{zhang2022dynamic,
  title={Dynamic Graph Message Passing Networks},
  author={Zhang, Li and Chen, Mohan and Arnab, Anurag and Xue, Xiangyang and Torr, Philip H.S.},
  journal={arXiv preprint arXiv:1908.06955},
  year={2022}
}

Acknowledgement

Thanks to previous open-sourced repo:
PVT
PyTorch Image Models
MMDetection
MMSegmentation