segformer

SegFormer

SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers

Introduction

Official Repo

Code Snippet

Abstract

We present SegFormer, a simple, efficient yet powerful semantic segmentation framework which unifies Transformers with lightweight multilayer perception (MLP) decoders. SegFormer has two appealing features: 1) SegFormer comprises a novel hierarchically structured Transformer encoder which outputs multiscale features. It does not need positional encoding, thereby avoiding the interpolation of positional codes which leads to decreased performance when the testing resolution differs from training. 2) SegFormer avoids complex decoders. The proposed MLP decoder aggregates information from different layers, and thus combining both local attention and global attention to render powerful representations. We show that this simple and lightweight design is the key to efficient segmentation on Transformers. We scale our approach up to obtain a series of models from SegFormer-B0 to SegFormer-B5, reaching significantly better performance and efficiency than previous counterparts. For example, SegFormer-B4 achieves 50.3% mIoU on ADE20K with 64M parameters, being 5x smaller and 2.2% better than the previous best method. Our best model, SegFormer-B5, achieves 84.0% mIoU on Cityscapes validation set and shows excellent zero-shot robustness on Cityscapes-C. Code will be released at: this http URL.

Citation

@article{xie2021segformer,
  title={SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers},
  author={Xie, Enze and Wang, Wenhai and Yu, Zhiding and Anandkumar, Anima and Alvarez, Jose M and Luo, Ping},
  journal={arXiv preprint arXiv:2105.15203},
  year={2021}
}

Usage

We have provided pretrained models converted from SegFormer.

If you want to convert keys on your own, we also provide a script mit2mmseg.py in the tools directory to convert the key of models from the official repo to MMSegmentation style.

python tools/model_converters/mit2mmseg.py ${PRETRAIN_PATH} ${STORE_PATH}

This script convert model from PRETRAIN_PATH and store the converted model in STORE_PATH.

Results and models

ADE20K

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	config	download
Segformer	MIT-B0	512x512	160000	2.1	38.17	37.85	38.97	config	model | log
Segformer	MIT-B1	512x512	160000	2.6	37.80	42.13	43.74	config	model | log
Segformer	MIT-B2	512x512	160000	3.6	26.80	46.80	48.12	config	model | log
Segformer	MIT-B3	512x512	160000	4.8	19.19	48.25	49.58	config	model | log
Segformer	MIT-B4	512x512	160000	6.1	14.54	49.09	50.72	config	model | log
Segformer	MIT-B5	512x512	160000	7.2	11.89	49.13	50.22	config	model | log
Segformer	MIT-B5	640x640	160000	11.5	10.60	50.19	51.41	config	model | log

Evaluation with AlignedResize:

Method	Backbone	Crop Size	Lr schd	mIoU	mIoU(ms+flip)
Segformer	MIT-B0	512x512	160000	38.55	39.03
Segformer	MIT-B1	512x512	160000	43.26	44.11
Segformer	MIT-B2	512x512	160000	47.46	48.16
Segformer	MIT-B3	512x512	160000	49.27	49.94
Segformer	MIT-B4	512x512	160000	50.23	51.10
Segformer	MIT-B5	512x512	160000	50.08	50.72
Segformer	MIT-B5	640x640	160000	51.13	51.66

Cityscapes

The lower fps result is caused by the sliding window inference scheme (window size:1024x1024).

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	mIoU	mIoU(ms+flip)	config	download
Segformer	MIT-B0	1024x1024	160000	3.64	4.74	76.54	78.22	config	model | log
Segformer	MIT-B1	1024x1024	160000	4.49	4.3	78.56	79.73	config	model | log
Segformer	MIT-B2	1024x1024	160000	7.42	3.36	81.08	82.18	config	model | log
Segformer	MIT-B3	1024x1024	160000	10.86	2.53	81.94	83.14	config	model | log
Segformer	MIT-B4	1024x1024	160000	15.07	1.88	81.89	83.38	config	model | log
Segformer	MIT-B5	1024x1024	160000	18.00	1.39	82.25	83.48	config	model | log

Note:

Original SegFormer paper uses different test_pipeline and image ratios in ms+flip. If you want to cite SegFormer original results as benchmark you may modify settings as below:

• We replace AlignedResize in original implementation to Resize + ResizeToMultiple. If you want to test by using AlignedResize, you can change the dataset pipeline like this:

test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(2048, 512),
        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            # resize image to multiple of 32, improve SegFormer by 0.5-1.0 mIoU.
            dict(type='ResizeToMultiple', size_divisor=32),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]

• Different from default setting of ms+flip, SegFormer original repo adopts different image ratios for ADE20K dataset. To re-implement numerical results of ms+flip, you can change image ratios in tools/test.py like this:

if args.aug_test:
    if cfg.data.test.type == 'ADE20KDataset':
        # hard code index
        cfg.data.test.pipeline[1].img_ratios = [
            0.75, 0.875, 1.0, 1.125, 1.25
        ]

• Training of SegFormer is not very stable, which is sensitive to random seeds.

• We use default training setting in MMSegmentation rather than RepeatDataset adopted in SegFormer official repo to accelerate training, here is its related issue.