Deep Hierarchical Semantic Segmentation

This repo contains the unoffical supported code and configuration files to reproduce semantic segmentaion results of HieraSeg. It is based on mmsegmentaion.

Updates

[2022-06-02] Initial commits

Results and Models

Dataset	Backbone	Crop Size	mIoU (single scale)	config	log	model
Cityscapes	Resnet101	512x1024	81.62	config	github	github
Pascal-Person-Part	ResNet101	480x480	73.44	config	github	github
LIP	ResNet101	480x480	58.71	config	github	github
Mapillary Vistas 2.0	ResNet101	512x1024	In progress	-	-	-

Usage

Installation

Please refer to get_started.md for installation and dataset preparation.

Requirement

Pytorch >= 1.8.0 & torchvision >= 0.9.0

Inference

# single-gpu testing
python tools/test.py <CONFIG_FILE> <SEG_CHECKPOINT_FILE> --eval mIoU

# multi-gpu testing
tools/dist_test.sh <CONFIG_FILE> <SEG_CHECKPOINT_FILE> <GPU_NUM> --eval mIoU

# multi-gpu, multi-scale testing
tools/dist_test.sh <CONFIG_FILE> <SEG_CHECKPOINT_FILE> <GPU_NUM> --aug-test --eval mIoU

Training

To train with pre-trained models, run:

# single-gpu training
python tools/train.py <CONFIG_FILE> --options model.pretrained=<PRETRAIN_MODEL> [model.backbone.use_checkpoint=True] [other optional arguments]

# multi-gpu training
tools/dist_train.sh <CONFIG_FILE> <GPU_NUM> --options model.pretrained=<PRETRAIN_MODEL> [model.backbone.use_checkpoint=True] [other optional arguments]

For example, to train on Cityscapes with a ResNet-101 backbone and 4 gpus, run:

tools/dist_train.sh configs/deeplabv3plus/deeplabv3plus_r101-d8_512x1024_80k_cityscapes_hiera_triplet.py.py 4

Notes:

We use four Tesla A100 GPUs for training. CUDA version: 11.1.

TODO

Code release
Checkpoint release
Implementation for Mapillary Vistats 2.0
HRNet and Swin-Transformer backbone

Citing HieraSeg

@article{li2022deep,
  title={Deep Hierarchical Semantic Segmentation},
  author={Li, Liulei and Zhou, Tianfei and Wang, Wenguan and Li, Jianwu and Yang, Yi},
  journal={CVPR},
  year={2022}
}

AI-Tianlong / HSSN_pytorch