Single Frame Semantic Segmentation Using Multi-Modal Spherical Images
The official implementation of Single Frame Semantic Segmentation Using Multi-Modal Spherical Images, Accepted to WACV 2024:
More details can be found in our paper [PDF].
<root>$ git clone git@github.com:sguttikon/SFSS-MMSI.git
<root>$ git clone -b sfss_mmsi git@github.com:sguttikon/matterport_utils.git
<root>$ git clone -b sfss_mmsi git@github.com:sguttikon/py360convert.git(base) <root>$ conda create --name sfss_mmsi
(base) <root>$ conda activate sfss_mmsi
(sfss_mmsi) <root>$ conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.7 -c pytorch -c nvidia
(sfss_mmsi) <root>$ conda install -c "nvidia/label/cuda-11.7.0" cuda-toolkit
(sfss_mmsi) <root>$ export CUDA_HOME=$CONDA_PREFIX
(sfss_mmsi) <root>$ python -m pip install timm==0.6.12 fvcore==0.1.5.post20221221 open3d==0.16.0 easydict==1.10 opencv-python==4.7.0.68 tensorboardx==2.5.1 notebook==7.0.2
(sfss_mmsi) <root>$ python -m pip install -U openmim==0.3.5
(sfss_mmsi) <root>$ mim install mmcv-full==1.6.2
(sfss_mmsi) <root>$ python -m pip install mmsegmentation==0.30.0
(sfss_mmsi) <root>$ python -m pip install -e py360convert
(sfss_mmsi) <root>$ cd SFSS-MMSI
(sfss_mmsi) <repo_path>$README.md
LICENSE
train.py
eval.py
/configs
/mp
/sct
/sid
/data_processing
/dataloader
/datasets
/2D-3D-Semantics-1K
/Structured3D-1K
/Matterport3D-1K
/engine
/models
/utils
/visualization
/pretrained
/segformers
/workdirs
/Stanford2D3DS_1024x512
/Structured3D_1024x512
/Matterport3D_1024x512
After installing the dependencies, we download and preprocess the datasets: Stanford2D3DS, Structured3D and Matterport3D. Refer to data_processing/.
Refer RGBX_Semantic_Segmentation for the pretrained segformer weights and place it under pretrained/segformers.
(sfss_mmsi) <repo_path>$ python -m train --config configs.sid.unimodal --devices 1
(sfss_mmsi) <repo_path>$ python -m train --config configs.sid.bimodal --devices 1
(sfss_mmsi) <repo_path>$ python -m train --config configs.sid.trimodal --devices 1
(sfss_mmsi) <repo_path>$ python -m train --config configs.sct.unimodal --devices 1
(sfss_mmsi) <repo_path>$ python -m train --config configs.sct.bimodal --devices 1
(sfss_mmsi) <repo_path>$ python -m train --config configs.sct.trimodal --devices 1(sfss_mmsi) <repo_path>$ python -m eval --config configs.sid.unimodal --split validation --epochs epoch-best.pth
(sfss_mmsi) <repo_path>$ python -m eval --config configs.sid.bimodal --split validation --epochs epoch-best.pth
(sfss_mmsi) <repo_path>$ python -m eval --config configs.sid.trimodal --split validation --epochs epoch-best.pth
(sfss_mmsi) <repo_path>$ python -m eval --config configs.sct.unimodal --split <validation/test> --epochs epoch-best.pth
(sfss_mmsi) <repo_path>$ python -m eval --config configs.sct.bimodal --split <validation/test> --epochs epoch-best.pth
(sfss_mmsi) <repo_path>$ python -m eval --config configs.sct.trimodal --split <validation/test> --epochs epoch-best.pth| Network | Fold | mAcc | aAcc | mIoU | Download |
|---|---|---|---|---|---|
| Unimodal (RGB) | avg | 63.955 | 80.417 | 52.873 | Fold1, Fold2, Fold3 |
| Bimodal (RGB-D) | avg | 66.021 | 84.889 | 55.492 | Fold1, Fold2, Fold3 |
| Bimodal (RGB-N) | avg | 68.787 | 86.686 | 58.239 | Fold1, Fold2, Fold3 |
| Bimodal (RGB-H) | avg | 70.683 | 88.066 | 60.603 | Fold1, Fold2, Fold3 |
| Trimodal (RGB-D-N) | avg | 69.031 | 87.322 | 59.426 | Fold1, Fold2, Fold3 |
| Network | valid mAcc | valid aAcc | valid mIoU | test mAcc | test aAcc | test mIoU | Download |
|---|---|---|---|---|---|---|---|
| Unimodal (RGB) | 80.476 | 95.729 | 71.941 | 76.099 | 95.388 | 68.343 | model |
| Bimodal (RGB-D) | 82.043 | 96.253 | 73.775 | 77.881 | 95.670 | 70.169 | model |
| Bimodal (RGB-N) | 83.247 | 96.500 | 74.378 | 78.683 | 96.115 | 71.001 | model |
| Trimodal (RGB-D-N) | 84.466 | 96.851 | 75.863 | 79.671 | 96.340 | 71.971 | model |
| Network | valid mAcc | valid aAcc | valid mIoU | test mAcc | test aAcc | test mIoU | Download |
|---|---|---|---|---|---|---|---|
| Unimodal (RGB) | 46.739 | 76.694 | 35.154 | 43.655 | 69.169 | 31.301 | model |
| Bimodal (RGB-D) | 51.395 | 78.763 | 39.191 | 49.242 | 71.592 | 35.921 | model |
| Bimodal (RGB-N) | 51.136 | 78.738 | 38.912 | 50.387 | 71.239 | 35.773 | model |
| Trimodal (RGB-D-N) | 51.563 | 78.937 | 39.263 | 50.007 | 71.422 | 35.520 | model |
In infer_pano_sem.ipynb, we use an out-of-training-distribution 360 image from PanoContext as an example.
Additionally, see infer_s2d3d_sem.ipynb, infer_s2d3d_sem.ipynb, and infer_mp3d_sem.ipynb for interactive demo and visualization.
If you are willing to contribute to this project, please contact with the authors.
Our code is heavily based on Trans4PASS+ and RGBX_Semantic_Segmentation, thanks for their excellent work!
We also appreciate the open-source works:
This work was partially funded by the EU Horizon Europe Framework Program under grant agreement 101058236 (HumanTech).
This repository is under the MIT license. For commercial use, please contact with the authors.
If you find this repo useful, please consider referencing the following paper:
@inproceedings{DBLP:conf/wacv/GuttikondaR24,
author = {Suresh Guttikonda and
Jason R. Rambach},
title = {Single Frame Semantic Segmentation Using Multi-Modal Spherical Images},
booktitle = {{WACV}},
pages = {3210--3219},
publisher = {{IEEE}},
year = {2024}
}
We are working on releasing the code and trained models, but no guarantee when the code will be released.
The code and trained models are released.