Squeeze-enhanced axial Transformer

Paper

SeaFormer: Squeeze-enhanced Axial Transformer for Mobile Semantic Segmentation,
Qiang Wan, Zilong Huang, Jiachen Lu, Gang Yu, Li Zhang
ICLR 2023

This repository contains the official implementation of SeaFormer.

SeaFormer achieves superior trade-off between performance and latency

The overall architecture of Seaformer

The schematic illustration of the SeaFormer layer

Model Zoo

Image Classification

Classification configs & weights see >>>here<<<.

• SeaFormer on ImageNet-1K

Model	Size	Acc@1	#Params (M)	FLOPs (G)
SeaFormer-Tiny	224	68.1	1.8	0.1
SeaFormer-Small	224	73.4	4.1	0.2
SeaFormer-Base	224	76.4	8.7	0.3
SeaFormer-Large	224	79.9	14.0	1.2

Semantic Segmentation

Segmentation configs & weights see >>>here<<<.

• SeaFormer on ADE20K

Method	Backbone	Pretrain	Iters	mIoU(ss)
Light Head	SeaFormer-Tiny	ImageNet-1K	160K	36.5
Light Head	SeaFormer-Small	ImageNet-1K	160K	39.4
Light Head	SeaFormer-Base	ImageNet-1K	160K	41.9
Light Head	SeaFormer-Large	ImageNet-1K	160K	43.8

• SeaFormer on Cityscapes

Method	Backbone	FLOPs	mIoU
Light Head(h)	SeaFormer-Small	2.0G	71.1
Light Head(f)	SeaFormer-Small	8.0G	76.4
Light Head(h)	SeaFormer-Base	3.4G	72.2
Light Head(f)	SeaFormer-Base	13.7G	77.7

BibTeX

@inproceedings{wan2023seaformer,
  title     = {SeaFormer: Squeeze-enhanced Axial Transformer for Mobile Semantic Segmentation},
  author    = {Wan, Qiang and Huang, Zilong and Lu, Jiachen and Yu, Gang and Zhang, Li},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2023}
}

Acknowledgment

Thanks to previous open-sourced repo:
TopFormer
mmsegmentation
pytorch-image-models