HILA: Improving Semantic Segmentation in Transformers using Hierarchical Inter-Level Attention

This is the official PyTorch implementation of HILA. For technical details, please refer to:

Improving Semantic Segmentation in Transformers using Hierarchical Inter-Level Attention.
Gary Leung, Jun Gao, Xiaohui Zeng, and Sanja Fidler.

University of Toronto Project page | Paper

Figure 1: Performance of adding HILA to SegFormer.

Figure 2: Adding HILA to pre-existing architectures.

Figure 3: Top-Down and Bottom-Up Inter-Level Attention.

License

MIT License

Copyright (c) 2022 Gary Leung, Jun Gao, Xiaohui Zeng, Sanja Fidler

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

Environment Setup

Clone this repo

git clone https://github.com/releasename
cd hila-master

Setup Python environment

conda create --name hila python=3.7.1
conda activate hila
pip install -r requirements.txt

Dataset setup

We use ADE20K and Cityscapes. For data preparation, please refer to MMSegmentation v0.13.0.

Models

Download trained weights and Imagenet-1K pretrained weights and put them in a folder pretrained/. Imagenet-1K pretrained models can be found in the pretrained/hila folder with model weights being in their respective folders. Segformer pretrained weights can be found here.

Evaluation

Example: evaluate SegFormer-B1 + HILA S(2,3,4) on Cityscapes:

# Single-scale testing
python ./tools/test.py local_configs/hila/segformer/B1/hila.b1.1024x1024.city.160k_S234.py  /path/to/checkpoint_file

# Multi-scale testing
python ./tools/test.py local_configs/hila/segformer/B1/hila.b1.1024x1024.city.160k_S234.py  /path/to/checkpoint_file --aug-test

# F-score testing
python ./tools/test.py local_configs/hila/segformer/B1/hila.b1.1024x1024.city.160k_S234.py  /path/to/checkpoint_file --eval-f1-start 0

# F-score testing (in batches due to memory issues)
# Note: mIOU code not integrated into batching, please run mIOU evaluation separately
python ./tools/test.py local_configs/hila/segformer/B1/hila.b1.1024x1024.city.160k_S234.py  /path/to/checkpoint_file --eval-f1-start 0 --eval-f1-step-size 100
python ./tools/test.py local_configs/hila/segformer/B1/hila.b1.1024x1024.city.160k_S234.py  /path/to/checkpoint_file --eval-f1-start 100 --eval-f1-step-size 100
...
python ./tools/test.py local_configs/hila/segformer/B1/hila.b1.1024x1024.city.160k_S234.py  /path/to/checkpoint_file --eval-f1-start 400 --eval-f1-step-size 100

# Modified F-score testing for ADE20K
python ./tools/test.py local_configs/hila/segformer/B1/hila.b1.512x512.ade.160k_S234.py  /path/to/checkpoint_file --eval-f1-start 0 --mod-f1

# Distance Crop testing (eg. 624 by 1248)
python ./tools/test.py local_configs/hila/segformer/B1/hila.b1.1024x1024.city.160k_S234.py  /path/to/checkpoint_file --eval-distance-crop 624 1248

Training

Example: train SegFormer-B1 + HILA S(2,3,4) on Cityscapes:

# GPU training
python ./tools/dist_train.sh local_configs/hila/segformer/B1/hila.b1.1024x1024.city.160k_S234.py <GPU_NUM>

Visualization

Example: visualize SegFormer-B1 + HILA S(2,3,4) on Cityscapes:

# Hierarchical visualization of Stage 4 feature at coor (x, y)
python ./tools/visualize_attention.py local_configs/hila/segformer/B1/hila.b1.1024x1024.city.160k_S234.py \
--show-dir ./path/output/dir/ --save-gt-seg --save-ori-img --data-range i j --attn-coors x y

Figure 4: Example Visualizations of Hierarchical Attention on Cityscapes.

Citation

@article{leung2022hila,
  title={HILA: Improving Semantic Segmentation in Transformers using Hierarchical Inter-Level Attention},
  author={Leung, Gary and Gao, Jun and Zeng, Xiaohui and Fidler, Sanja},
  journal={arXiv:2207.02126},
  year={2022}
}

fidler-lab/hila