Fabio Cermelli, Matthieu Cord, Arthur Douillard

[ arXiv ] [ BibTeX ]

See installation instructions.

See Preparing Datasets for Mask2Former.

Per-Pixel baseline: MODEL.MASK_FORMER.PER_PIXEL True

Mask-based methods: MODEL.MASK_FORMER.SOFTMASK True MODEL.MASK_FORMER.FOCAL True

CoMFormer: CONT.DIST.PSEUDO True CONT.DIST.KD_WEIGHT 10.0 CONT.DIST.UKD True CONT.DIST.KD_REW True

MiB: CONT.DIST.KD_WEIGHT 200.0 CONT.DIST.UKD True CONT.DIST.UCE True

PLOP: CONT.DIST.PSEUDO True CONT.DIST.PSEUDO_TYPE 1 CONT.DIST.POD_WEIGHT 0.001

ADE Semantic Segmenation:

• Use config file: cfg_file=configs/ade20k/semantic-segmentation/maskformer2_R101_bs16_90k.yaml
• 100-50: CONT.BASE_CLS 100 CONT.INC_CLS 50 CONT.MODE overlap (see examples in scripts/ade.sh)
• 100-10: CONT.BASE_CLS 100 CONT.INC_CLS 10 CONT.MODE overlap (see examples in scripts/ade10.sh)
• 100-5: CONT.BASE_CLS 100 CONT.INC_CLS 5 CONT.MODE overlap (see examples in scripts/ade5.sh)

ADE Panoptic Segmenation:

• Use config file: cfg_file=configs/ade20k/panoptic-segmentation/maskformer2_R50_bs16_90k.yaml
• 100-50: CONT.BASE_CLS 100 CONT.INC_CLS 50 CONT.MODE overlap (see examples in scripts/adps.sh)
• 100-10: CONT.BASE_CLS 100 CONT.INC_CLS 10 CONT.MODE overlap (see examples in scripts/adps10.sh)
• 100-5: CONT.BASE_CLS 100 CONT.INC_CLS 5 CONT.MODE overlap (see examples in scripts/adps5.sh)

If you use CoMFormer in your research, please use the following BibTeX entry.

@article{cermelli2023comformer,
  title={CoMFormer: Continual Learning in Semantic and Panoptic Segmentation},
  author={Fabio Cermelli and Matthieu Cord and Arthur Douillard},
  journal={IEEE/CVF Computer Vision and Pattern Recognition Conference},
  year={2023}
}

The code is largely based on Mask2Former.