A³-FPN: Asymptotic Content-Aware Pyramid Attention Network for Dense Visual Prediction

This repository is the implementation for our paper "A³-FPN: Asymptotic Content-Aware Pyramid Attention Network for Dense Visual Prediction". A³-FPN employs a horizontally-spread column network that enables asymptotically global feature interaction and disentangles each level from all hierarchical representations. In feature fusion, it collects supplementary content from the adjacent level to generate position-wise offsets and weights for context-aware resampling, and learns deep context reweights to improve intra-category similarity. In feature reassembly, it further strengthens intra-scale discriminative feature learning and reassembles redundant features based on information content and spatial variation of feature maps. Extensive experiments on MS COCO, VisDrone2019-DET and Cityscapes demonstrate that A³-FPN can be easily integrated into state-of-the-art CNN and Transformer-based architectures, yielding remarkable performance gains.

Install

This project is based on mmdetection, mmsegmentation and detectron2. Please refer to mmdetection, mmsegmentation, detectron2, Mask2Former and dcnv4 for constructing the running environment.

Train

Single gpu for training:

CUDA_VISIBLE_DEVICES=0 python ./mmdetection/tools/train.py <config.yaml> --work-dir ./weight

python ./detectron2/tools/train_net.py --config-file <config.yaml> --num-gpus 1

Multiple gpus for training:

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ./mmdetection/tools/dist_train.sh <config.yaml> 8 --work-dir ./weight

python3 ./detectron2/tools/train_net.py --config-file <config.yaml> --num-gpus 8

If you want to train more models, please refer to train.py.

Test / Evaluate

CUDA_VISIBLE_DEVICES=0 python ./mmdetection/tools/test.py <config.yaml> <CHECKPOINT_FILE>

python3 ./detectron2/tools/train_net.py --config-file <config.yaml> --num-gpus 1 --eval-only MODEL.WEIGHTS /path/to/model_checkpoint

If you want to test more models, please refer to test.py.

Results

PASCAL VOC Object Detection

Model	Backbone	Lr Sched	AP50	AP75	Weight
Faster R-CNN + A3-FPN	ResNet-50	18k	82.63	62.55	Model

VisDrone2019-DET Object Detection

Model	Backbone	Lr Sched	AP	AP50	AP75	Weight
RetinaNet + A3-FPN	ResNet-50	1x	23.7	39.4	24.7	Model

COCO Object Detection and Instance Segmentation

Model	Backbone	Lr Sched	APbox	APmask	Weight
Mask R-CNN + A3-FPN	ResNet-50	3x	43.70	39.19	Model
Mask2Former + A3-FPN	ResNet-50	50 Epoch	-	44.28	Model

Cityscapes Semantic Segmentation

Model	Backbone	Lr Sched	mIoUs.s.	mIoUm.s.	Weight
UperNet + A3-FPN	ResNet-50	80k	79.65	81.22	Model
Mask2Former + A3-FPN	ResNet-50	90k	81.13	-	Model

Citation

If you find A³-FPN useful in your research, please consider citing:

@article{qin2026a3,
  title={A3-FPN: Asymptotic Content-Aware Pyramid Attention Network for Dense Visual Prediction},
  author={Qin, Meng'en and Song, Yu and Zhao, Quanling and Yang, Xiaodong and Che, Yingtao and Yang, Xiaohui},
  journal={arXiv preprint arXiv:2604.10210},
  year={2026}
}