Segmenter: Transformer for Semantic Segmentation

Segmenter: Transformer for Semantic Segmentation by Robin Strudel*, Ricardo Garcia*, Ivan Laptev and Cordelia Schmid.

*Equal Contribution

Installation

Define os environment variables pointing to your checkpoint and dataset directory, put in your .bashrc:

export DATASET=/path/to/dataset/dir

Install PyTorch 1.9 then pip install . at the root of this repository.

To download ADE20K, use the following command:

python -m segm.scripts.prepare_ade20k $DATASET

Model Zoo

We release models with a Vision Transformer backbone initialized from the improved ViT models.

ADE20K

Segmenter models with ViT backbone:

Name	mIoU (SS/MS)	# params	Resolution	FPS	Download
Seg-T-Mask/16	38.1 / 38.8	7M	512x512	52.4	model	config	log
Seg-S-Mask/16	45.3 / 46.9	27M	512x512	34.8	model	config	log
Seg-B-Mask/16	48.5 / 50.0	106M	512x512	24.1	model	config	log
Seg-B/8	49.5 / 50.5	89M	512x512	4.2	model	config	log
Seg-L-Mask/16	51.8 / 53.6	334M	640x640	-	model	config	log

Segmenter models with DeiT backbone:

Name	mIoU (SS/MS)	# params	Resolution	FPS	Download
Seg-B†/16	47.1 / 48.1	87M	512x512	27.3	model	config	log
Seg-B†-Mask/16	48.7 / 50.1	106M	512x512	24.1	model	config	log

Pascal Context

Name	mIoU (SS/MS)	# params	Resolution	FPS	Download
Seg-L-Mask/16	58.1 / 59.0	334M	480x480	-	model	config	log

Inference

Download one checkpoint with its configuration in a common folder, for example seg_tiny_mask.

You can generate segmentation maps from your own data with:

python -m segm.inference --model-path seg_tiny_mask/checkpoint.pth -i images/ -o segmaps/

To evaluate on ADE20K, run the command:

# single-scale evaluation:
python -m segm.eval.miou seg_tiny_mask/checkpoint.pth ade20k --singlescale
# multi-scale evaluation:
python -m segm.eval.miou seg_tiny_mask/checkpoint.pth ade20k --multiscale

Train

Train Seg-T-Mask/16 on ADE20K on a single GPU:

python -m segm.train --log-dir seg_tiny_mask --dataset ade20k \
  --backbone vit_tiny_patch16_384 --decoder mask_transformer

To train Seg-B-Mask/16, simply set vit_base_patch16_384 as backbone and launch the above command using a minimum of 4 V100 GPUs (~12 minutes per epoch) and up to 8 V100 GPUs (~7 minutes per epoch). The code uses SLURM environment variables.

Logs

To plot the logs of your experiments, you can use

python -m segm.utils.logs logs.yml

with logs.yml located in utils/ with the path to your experiments logs:

root: /path/to/checkpoints/
logs:
  seg-t: seg_tiny_mask/log.txt
  seg-b: seg_base_mask/log.txt

Attention Maps

To visualize the attention maps for Seg-T-Mask/16 encoder layer 0 and patch (0, 21), you can use:

python -m segm.scripts.show_attn_map seg_tiny_mask/checkpoint.pth \ 
images/im0.jpg output_dir/ --layer-id 0 --x-patch 0 --y-patch 21 --enc

Different options are provided to select the generated attention maps:

--enc\--dec: Select encoder or decoder attention maps respectively.
--patch\--cls: Patch mode (--patch) generates attention maps for the patch provided with flags --x-patch and --y-patch while CLS mode (--cls) generate attention maps for the CLS token in the encoder if --enc flag is provided and for each class embedding in the decoder if --dec is used.
--x-patch and --y-patch: Select the patch for which attention maps are generated. This flag is ignored when --cls flag is used.
--layer-id: Select the layer for which the attention maps are generated.

For example, to generate attention maps for the decoder class embeddings, you can use:

python -m segm.scripts.show_attn_map seg_tiny_mask/checkpoint.pth \
images/im0.jpg output_dir/ --layer-id 0 --dec --cls

Attention maps for patch (0, 21) in Seg-L-Mask/16 encoder layers 1, 4, 8, 12 and 16:

Attention maps for the class embeddings in Seg-L-Mask/16 decoder layer 0:

Video Segmentation

Zero shot video segmentation on DAVIS video dataset with Seg-B-Mask/16 model trained on ADE20K.

BibTex

@article{strudel2021,
  title={Segmenter: Transformer for Semantic Segmentation},
  author={Strudel, Robin and Garcia, Ricardo and Laptev, Ivan and Schmid, Cordelia},
  journal={arXiv preprint arXiv:2105.05633},
  year={2021}
}

Acknowledgements

The Vision Transformer code is based on timm library and the semantic segmentation training and evaluation pipeline is using mmsegmentation.

rjgpinel/segmenter