Skip to content

Latest commit

 

History

History
114 lines (75 loc) · 4.75 KB

README.md

File metadata and controls

114 lines (75 loc) · 4.75 KB

Open-set Semantic Segmentation Built atop Mask-level Recognition

This repository provides the official implementation of our recent papers:

📜 UNO: Outlier detection by ensembling uncertainty with negative objectness

Anja Delić, Matej Grcić, Siniša Šegvić

Published in BMVC 2024

[arXiv]

📜 EAM: On Advantages of Mask-level Recognition for Outlier-aware Segmentation

Matej Grcić, Josip Šarić, Siniša Šegvić

Published in CVPR workshop (VAND) 2023

[arXiv]

Installation

Similar to Mask2Former repo, see installation instructions.

Training

UNO

Finetuning a model with the additional negative objectness class with ADE20K negatives:

python finetune_UNO.py --config-file configs/cityscapes/semantic-segmentation/swin/maskformer2_swin_large_IN21k_384_bs18_2k_city+vistas_uno.yaml --num-gpus 3

Finetuning a model with the additional negative objectness class with synthetic negatives:

python finetune_UNO_synthetic.py --config-file  configs/cityscapes/semantic-segmentation/swin/maskformer2_swin_large_IN21k_384_bs18_2k_city+vistas_uno_synthetic.yaml --num-gpus 3

EAM

Rejecting predictions in negative instances:

python train_net.py --config-file configs/cityscapes/semantic-segmentation/swin/maskformer2_swin_large_IN21k_384_bs12_2k_city+vistas_oe.yaml

Evaluation

UNO

python train_net.py --config-file  configs/cityscapes/semantic-segmentation/swin/maskformer2_swin_large_IN21k_384_bs18_115k_city+_vistas_uno.yaml --eval-only MODEL.WEIGHTS path_to_model DATASETS.TEST eval_dataset_name

EAM

python train_net.py --config-file  configs/cityscapes/semantic-segmentation/swin/maskformer2_swin_large_IN21k_384_bs18_115k_city+vistas.yaml --eval-only MODEL.WEIGHTS path_to_model DATASETS.TEST eval_dataset_name

eval_dataset_name can be one of the following: ("fs_static_val", "fs_laf_val", "road_anomaly",)

Checkpoints

EAM

Mask2Former with SWIN-L backbone trained on Cityscapes (CS): weights

Mask2Former with SWIN-L backbone trained on Cityscapes and Vistas (CS&MV): weights

Mask2Former with SWIN-L backbone (CS&MV) fine-tuned with ADE20k negatives: weights

UNO

Mask2Former with SWIN-L backbone (CS&MV) with K+2 classes fine-tuned with ADE20k negatives: weights

Mask2Former with SWIN-L backbone (CS&MV) with K+2 classes fine-tuned with synthetic negatives: weights
DenseFlow pretrained on CS&MV: weights

License

Shield: License: MIT

The majority of Mask2Former is licensed under a MIT License.

However portions of the project are available under separate license terms: Swin-Transformer-Semantic-Segmentation is licensed under the MIT license, Deformable-DETR is licensed under the Apache-2.0 License.

Citing Outlier-aware Mask2Former

@inproceedings{delic24bmvc,
  title={Outlier detection by ensembling uncertainty with negative objectness},
  author={Anja Delić and Matej Grcic and Siniša Šegvić}
  journal={BMVC 2024 British Machine Vision Conference},
  year={2024}
}
@inproceedings{grcic23cvprw,
  title={On Advantages of Mask-level Recognition for Outlier-aware Segmentation},
  author={Matej Grcic and Josip Šarić and Siniša Šegvić}
  journal={CVPR 2023 Workshop Visual Anomaly and Novelty Detection (VAND)},
  year={2023}
}
@inproceedings{cheng2021mask2former,
  title={Masked-attention Mask Transformer for Universal Image Segmentation},
  author={Bowen Cheng and Ishan Misra and Alexander G. Schwing and Alexander Kirillov and Rohit Girdhar},
  journal={CVPR},
  year={2022}
}

Acknowledgement

Code is extension of Mask2Former (https://github.com/facebookresearch/Mask2Former).