swjtulinxi / QuadTreeAttention

QuadTree Attention for Vision Transformers (ICLR2022)

This repository contains codes for quadtree attention. This repo contains codes for feature matching, image classficiation, object detection and semantic segmentation.

Installation

1. Compile the quadtree attention operation cd QuadTreeAttention&&python setup.py install
2. Install the package for each task according to each README.md in the separate directory.

Model Zoo and Baselines

We provide baselines results and model zoo in the following.

Feature matching

News! QuadTree Attention achieves the best single model performance among all public available pretrained models in image matching chanllenge 2022. Please refer to this [post].

• Quadtree on Feature matching

Method	AUC@5	AUC@10	AUC@20	Model
ScanNet	24.9	44.7	61.8	[Google]/[GitHub]
Megadepth	53.5	70.2	82.2	[Google]/[GitHub]

Image classification

• Quadtree on ImageNet-1K

Method	Flops	Acc@1	Model
Quadtree-B-b0	0.6	72.0	[Google]/[GitHub]
Quadtree-B-b1	2.3	80.0	[Google]/[GitHub]
Quadtree-B-b2	4.5	82.7	[Google]/[GitHub]
Quadtree-B-b3	7.8	83.8	[Google]/[GitHub]
Quadtree-B-b4	11.5	84.0	[Google]/[GitHub]

Object detection and instance segmentation

• Quadtree on COCO

Baseline Detectors

Method	Backbone	Pretrain	Lr schd	Aug	Box AP	Mask AP	Model
RetinaNet	Quadtree-B-b0	ImageNet-1K	1x	No	38.4	-	[Google]/[GitHub]
RetinaNet	Quadtree-B-b1	ImageNet-1K	1x	No	42.6	-	[Google]/[GitHub]
RetinaNet	Quadtree-B-b2	ImageNet-1K	1x	No	46.2	-	[Google]/[GitHub]
RetinaNet	Quadtree-B-b3	ImageNet-1K	1x	No	47.3	-	[Google]/[GitHub]
RetinaNet	Quadtree-B-b4	ImageNet-1K	1x	No	47.9	-	[Google]/[GitHub]
Mask R-CNN	Quadtree-B-b0	ImageNet-1K	1x	No	38.8	36.5	[Google]/[GitHub]
Mask R-CNN	Quadtree-B-b1	ImageNet-1K	1x	No	43.5	40.1	[Google]/[GitHub]
Mask R-CNN	Quadtree-B-b2	ImageNet-1K	1x	No	46.7	42.4	[Google]/[GitHub]
Mask R-CNN	Quadtree-B-b3	ImageNet-1K	1x	No	48.3	43.3	[Google]/[GitHub]
Mask R-CNN	Quadtree-B-b4	ImageNet-1K	1x	No	48.6	43.6	[Google]/[GitHub]

Semantic Segmentation

• Quadtree on ADE20K

Method	Backbone	Pretrain	Iters	mIoU	Model
Semantic FPN	Quadtree-b0	ImageNet-1K	160K	39.9	[Google]/[GitHub]
Semantic FPN	Quadtree-b1	ImageNet-1K	160K	44.7	[Google]/[GitHub]
Semantic FPN	Quadtree-b2	ImageNet-1K	160K	48.7	[Google]/[GitHub]
Semantic FPN	Quadtree-b3	ImageNet-1K	160K	50.0	[Google]/[GitHub]
Semantic FPN	Quadtree-b4	ImageNet-1K	160K	50.6	[Google]/[GitHub]

Citation

@article{tang2022quadtree,
  title={QuadTree Attention for Vision Transformers},
  author={Tang, Shitao and Zhang, Jiahui and Zhu, Siyu and Tan, Ping},
  journal={ICLR},
  year={2022}
}