MaskFormer: Per-Pixel Classification is Not All You Need for Semantic Segmentation

Official Repo

Code Snippet

Modern approaches typically formulate semantic segmentation as a per-pixel classification task, while instance-level segmentation is handled with an alternative mask classification. Our key insight: mask classification is sufficiently general to solve both semantic- and instance-level segmentation tasks in a unified manner using the exact same model, loss, and training procedure. Following this observation, we propose MaskFormer, a simple mask classification model which predicts a set of binary masks, each associated with a single global class label prediction. Overall, the proposed mask classification-based method simplifies the landscape of effective approaches to semantic and panoptic segmentation tasks and shows excellent empirical results. In particular, we observe that MaskFormer outperforms per-pixel classification baselines when the number of classes is large. Our mask classification-based method outperforms both current state-of-the-art semantic (55.6 mIoU on ADE20K) and panoptic segmentation (52.7 PQ on COCO) models.

• MaskFormer model needs to install MMDetection first.

pip install "mmdet>=3.0.0rc4"

Note:

• All experiments of MaskFormer are implemented with 8 V100 (32G) GPUs with 2 samplers per GPU.
• The results of MaskFormer are relatively not stable. The accuracy (mIoU) of model with R-101-D32 is from 44.7 to 46.0, and with Swin-S is from 49.0 to 49.8.
• The ResNet backbones utilized in MaskFormer models are standard ResNet rather than ResNetV1c.
• Test time augmentation is not supported in MMSegmentation 1.x version yet, we would add "ms+flip" results as soon as possible.

@article{cheng2021per,
  title={Per-pixel classification is not all you need for semantic segmentation},
  author={Cheng, Bowen and Schwing, Alex and Kirillov, Alexander},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  pages={17864--17875},
  year={2021}
}