aim-uofa / NRD_decoder

Dynamic Neural Representational Decoders for High-Resolution Semantic Segmentation

Dynamic Neural Representational Decoders for High-Resolution Semantic Segmentation, NeurIPS'21

arXiv paper or the final published version

Requirements

This repository needs mmsegmentation

Training

To train the model(s) in the paper, run this command:

python tools/train.py ./configs/NRD/ade20k/NRD_r101_512x512_164k_ade20k.py

The batch size is 16 in this work. Please change the 'samples_per_gpu' in configs/base/datasets/.. accordingly

Evaluation

To evaluate my model at single-scale inference, run:

python tools/eval.py ./configs/NRD/ade20k/NRD_r101_512x512_164k_ade20k.py  {path-to-checkpoint-file}   --eval mIoU

Pre-trained Models

Results

Our model achieves the following performance on :

[Semantic segmentation results]

Model name	datasets	mIoU	mIoU (ms)
NRD-r101	ade20k (val)	44.01	45.62
NRD-x101	ade20k (val)	44.34	46.35
NRD-r101	pascal-context(val)	52.31 (59 classes)	54.1 (59 classes)
NRD-r101	pascal-context(val)	47.5 (60 classes)	40.9 (60 classes)
NRD-r50	Cityscapes (val)	79.8	80.8
NRD-r101	Cityscapes (val)	80.7	82.0

BibTeX

@inproceedings{NEURIPS2021_Zhangbowen,
 author = {Bowen Zhang and Yifan liu and Zhi Tian and Chunhua Shen},
 booktitle = {Proc.\ Advances in Neural Information Processing Systems},
 title = {Dynamic Neural Representational Decoders for High-Resolution Semantic Segmentation},
 year = {2021}
}

Contributing

This code is built upon mmsegmentation.