Channelized Axial Attention for Semantic Segmentation (AAAI-2022)

News

Mar-16-2022: Based on the latest Class-aware Regularizations (CAR), CAA + ConvNeXt-Large achieved 64.12% mIOU on Pascal Context! The repo of CAR also contains the train code for CAA.

Some segmentation results on Flickr images:

Installation

1. Install TensorFlow (>= 2.4, 2.3 is not recommend for GPU, but okay for TPU)
2. Install iSeg (My personal segmentation codebase, update soon at https://github.com/edwardyehuang/iSeg)
3. Install iDS (Dataset supports for iSeg, update soon at https://github.com/edwardyehuang/iDS)
4. Clone this repo

Model Zoo

Since some of the original experiments (especially for ResNet-101) are conducted a long time ago, the ckpts listed below may have slightly different performance with paper reported.

Pascal Context

Backbone	ckpts	mIOU%	configs
ResNet-101	weiyun	55.0	configs
EfficientNet-B7	weiyun	60.3	configs

COCOStuff-10k

Backbone	ckpts	mIOU%	configs
ResNet-101	weiyun	41.2	configs

COCOStuff-164k

Backbone	ckpts	mIOU%	configs
EfficientNet-B5	weiyun	47.27	configs

Please cite us

@InProceedings{cCAA,
	author = "Ye Huang and Di Kang and Wenjing Jia and Xiangjian He and Liu liu",
	title = "Channelized Axial Attention - Considering Channel Relation within Spatial Attention for Semantic Segmentation",
	booktitle = "Proceedings of the AAAI Conference on Artificial Intelligence",
	year = "2022",
	DOI={10.1609/aaai.v36i1.19985},
}