DGMN2
DGMN2 copied to clipboard
Published
20 hours ago •
fudan-zvg
fudan-zvg
[TPAMI 2022 & CVPR 2020 Oral] Dynamic Graph Message Passing Networks
DGMN2
This repository contains the implementation of Dynamic Graph Message Passing Networks for Visual Recognition.
Main results
Image Classification
ImageNet-1K
| Method | Params (M) | FLOPs (G) | Top1 Acc (%) | Download |
|---|---|---|---|---|
| DGMN2-Tiny | 12.1 | 2.3 | 78.7 | model |
| DGMN2-Small | 21.0 | 4.3 | 81.7 | model |
| DGMN2-Medium | 35.8 | 7.1 | 82.5 | model |
| DGMN2-Large | 48.3 | 10.4 | 83.3 | model |
Object Detection
COCO validation set
| Method | Backbone | Lr schd | box AP | mask AP | Download |
|---|---|---|---|---|---|
| RetinaNet | DGMN2-Tiny | 1x | 39.7 | - | model |
| RetinaNet | DGMN2-Small | 1x | 42.5 | - | model |
| RetinaNet | DGMN2-Medium | 1x | 43.7 | - | model |
| RetinaNet | DGMN2-Large | 1x | 44.7 | - | model |
| Mask R-CNN | DGMN2-Tiny | 1x | 40.2 | 37.2 | model |
| Mask R-CNN | DGMN2-Small | 1x | 43.4 | 39.7 | model |
| Mask R-CNN | DGMN2-Medium | 1x | 44.4 | 40.2 | model |
| Mask R-CNN | DGMN2-Large | 1x | 46.2 | 41.6 | model |
| Deformable DETR | DGMN2-Tiny | 50e | 44.4 | - | model |
| Deformable DETR | DGMN2-Small | 50e | 47.3 | - | model |
| Deformable DETR | DGMN2-Medium | 50e | 48.4 | - | model |
| Deformable DETR+ | DGMN2-Small | 50e | 48.5 | - | model |
| Sparse R-CNN | DGMN2-Small | 3x | 48.2 | - | model |
Semantic Segmentation
Cityscapes validation set
| Method | Backbone | Iters | mIoU | mIoU (ms + flip) | Download |
|---|---|---|---|---|---|
| Semantic FPN | DGMN2-Tiny | 40K | 78.09 | 79.40 | model |
| Semantic FPN | DGMN2-Small | 40K | 80.65 | 81.58 | model |
| Semantic FPN | DGMN2-Medium | 40K | 80.60 | 81.79 | model |
| Semantic FPN | DGMN2-Large | 40K | 81.75 | 82.64 | model |
| SETR-Naive | DGMN2-Tiny | 40K | 77.23 | 78.23 | model |
| SETR-Naive | DGMN2-Small | 40K | 80.31 | 81.04 | model |
| SETR-Naive | DGMN2-Medium | 40K | 80.83 | 81.39 | model |
| SETR-Naive | DGMN2-Large | 40K | 81.80 | 82.61 | model |
| SETR-PUP | DGMN2-Tiny | 40K | 78.25 | 79.26 | model |
| SETR-PUP | DGMN2-Small | 40K | 79.78 | 80.73 | model |
| SETR-PUP | DGMN2-Medium | 40K | 80.97 | 81.80 | model |
| SETR-PUP | DGMN2-Large | 40K | 81.58 | 82.27 | model |
| SETR-MLA | DGMN2-Tiny | 40K | 78.25 | 79.32 | model |
| SETR-MLA | DGMN2-Small | 40K | 80.79 | 81.62 | model |
| SETR-MLA | DGMN2-Medium | 40K | 81.09 | 82.00 | model |
| SETR-MLA | DGMN2-Large | 40K | 81.55 | 81.98 | model |
Getting Started
- For image classification, please see classification.
- For object detection, please see detection.
- For semantic segmentation, please see segmentation.
- For Deformable DETR, please see deformable_detr.
- For Sparse R-CNN, please see sparsercnn.
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 for Visual Recognition},
author={Zhang, Li and Chen, Mohan and Arnab, Anurag and Xue, Xiangyang and Torr, Philip H.S.},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
year={2022}
}
Acknowledgement
Thanks to previous open-sourced repo:
PVT
PyTorch Image Models
MMDetection
MMSegmentation
Deformable DETR
Sparse R-CNN
Detectron2
fudan-zvg