WeSAM

Improving the Generalization of Segmentation Foundation Model under Distribution Shift via Weakly Supervised Adaptation

CVPR 2024

Haojie Zhang¹*, Yongyi Su¹*, Xun Xu²⁺, Kui Jia³

¹South China University of Technology ²Institute for Infocomm Research, A*STAR
³School of Data Science, The Chinese University of Hong Kong, Shenzhen

Paper

Code

Abstract

The success of large language models has inspired the computer vision community to explore image segmentation foundation model that is able to zero/few-shot generalize through prompt engineering. Segment-Anything~(SAM), among others, is the state-of-the-art image segmentation foundation model demonstrating strong zero/few-shot generalization. Despite the success, recent studies reveal the weakness of SAM under strong distribution shift. In particular, SAM performs awkwardly on corrupted natural images, camouflaged images, medical images, etc. Motivated by the observations, we aim to develop a self-training based strategy to adapt SAM to target distribution. Given the unique challenges of large source dataset, high computation cost and incorrect pseudo label, we propose a weakly supervised self-training architecture with anchor regularization and low-rank finetuning to improve the robustness and computation efficiency of adaptation. We validate the effectiveness on 5 types of downstream segmentation tasks including natural clean/corrupted images, medical images, camouflaged images and robotic images. Our proposed method is task-agnostic in nature and outperforms pre-trained SAM and state-of-the-art domain adaptation methods on almost all downstream tasks with the same testing prompt inputs.

Contributions & Method

We summarize the contributions of this work as follows:

We are motivated by the generalization issue of segment-anything (SAM) model to diverse downstream segmentation tasks and propose a task-agnostic solution to adapt SAM through self-training with no access to source dataset.
We exploit weak supervisions, including bounding box, point-wise annotation and coarse segmentation masks, to improve the effectiveness of adaptation. The weak supervision are fully compatible with the prompt encoder of SAM.
Extensive experiments on 5 types of downstream instance segmentation tasks demonstrate the effectiveness of the proposed weakly supervised adaptation approach.

The proposed self-training architecture with anchor network regularization and contrastive loss regularization. Red arrows indicates the backpropagation flow.

Quantitative Evaluations

Table 1: Adaptation results on COCO-C dataset using bounding box prompt.

Method	Brit	Contr	Defoc	Elast	Fog	Frost	Gauss	Glass	Impul	Jpeg	Motn	Pixel	Shot	Snow	Zoom	Avg
Direct	72.83	57.34	64.47	69.36	72.39	70.50	67.20	64.43	67.65	68.23	62.72	68.60	67.44	69.02	58.80	66.73
TENT	76.02	61.51	67.48	70.88	74.89	73.88	69.01	67.10	69.28	70.25	65.45	70.81	69.96	72.37	62.59	69.43
SHOT	73.84	59.09	65.91	69.57	73.98	72.51	68.30	66.09	68.61	69.45	64.56	70.48	68.77	71.03	60.17	68.16
Soft Teacher	73.90	62.12	65.41	71.32	72.16	73.27	68.84	67.49	68.73	70.18	66.88	69.79	70.08	73.33	64.88	69.23
TRIBE	76.40	60.86	66.19	72.72	75.08	75.14	70.34	66.66	70.83	72.42	65.94	70.24	70.66	74.22	64.56	70.15
DePT	69.15	57.26	59.08	66.80	58.73	66.75	66.78	62.74	65.65	66.39	61.66	66.65	67.57	66.62	58.21	64.42
WDASS	76.21	60.57	67.07	72.34	75.97	74.63	69.84	67.88	69.92	71.36	66.25	71.99	70.32	72.25	63.61	70.01
OURS	78.50	61.05	66.99	73.93	77.09	76.10	72.02	68.21	71.29	72.77	66.33	70.90	70.28	75.07	65.33	71.05
Supervised	78.86	74.81	72.04	74.32	78.01	77.14	73.43	72.12	74.08	75.30	71.39	75.15	74.25	76.34	68.04	74.35

Table 2: Adaptation results on natural clean image datasets.

Method	COCO 2017			Pascal VOC
Method	box	point	poly	box	point	poly
Direct	74.29	55.06	65.64	69.21	69.21	60.79
TENT	78.21	52.99	71.51	80.24	74.97	65.03
SHOT	75.18	58.46	69.26	79.80	74.26	63.38
Soft Teacher	75.94	43.36	68.27	72.93	56.09	62.20
TRIBE	77.56	49.56	70.99	78.87	69.21	65.39
DePT	71.00	37.35	63.27	74.09	42.99	59.94
WDASS	77.29	60.55	70.19	80.12	76.15	66.98
OURS	80.12	62.09	72.33	80.27	74.15	66.72
Supervised	81.50	69.77	73.39	81.23	76.98	71.32

Table 3: Adaptation results on medical image segmentation datasets.

Method	kvasir-SEG			ISIC
Method	box	point	poly	box	point	poly
Direct	81.59	62.30	54.03	66.74	53.42	62.82
TENT	82.47	61.84	62.97	71.76	53.46	67.12
SHOT	82.30	63.76	61.34	71.99	55.99	66.86
Soft Teacher	84.12	73.53	58.15	75.74	54.95	72.29
TRIBE	85.05	73.03	64.61	72.61	50.36	67.99
DePT	81.91	52.06	61.55	78.43	46.79	72.75
WDASS	84.01	63.78	64.78	74.23	55.63	67.84
OURS	85.47	75.23	67.40	80.01	62.12	75.36
Supervised	85.89	77.54	81.64	81.62	79.81	80.26

Table 4: Adaptation results on camouflaged object datasets.

Method	CHAMELEON			CAMO			COD10K
Method	box	point	poly	box	point	poly	box	point	poly
Direct	51.32	39.37	45.78	62.72	57.43	50.85	66.32	63.61	40.04
TENT	65.48	54.53	53.06	71.24	59.59	60.29	69.36	61.94	43.36
SHOT	68.60	62.47	54.36	71.61	62.78	58.72	69.09	65.25	42.38
Soft Teacher	65.92	44.17	46.72	62.30	48.64	51.26	66.32	50.04	32.27
TRIBE	71.00	52.80	54.99	66.00	61.97	60.54	67.84	63.62	42.75
DePT	54.48	33.46	42.47	55.44	33.07	48.63	59.32	34.06	35.51
WDASS	71.91	62.40	56.80	71.25	63.39	62.29	71.42	65.61	43.93
OURS	75.94	74.00	66.83	73.42	65.55	62.90	71.93	70.55	45.87
Supervised	78.05	85.86	68.38	79.17	77.01	67.12	78.06	78.44	64.90

Table 5: Adaptation results on robotic image datasets.

Method	OCID			OSD
Method	box	point	poly	box	point	poly
Direct	86.35	71.41	72.81	87.62	78.86	80.77
TENT	87.77	66.61	77.53	88.10	80.53	87.85
SHOT	88.06	74.39	76.25	88.09	80.52	87.86
Soft Teacher	84.98	68.46	73.75	90.41	80.49	87.00
TRIBE	86.77	67.86	76.50	90.42	80.54	87.84
DePT	82.00	56.52	70.92	81.84	69.06	82.50
WDASS	87.68	77.13	76.70	88.07	80.52	88.19
OURS	88.09	80.14	77.41	92.11	80.51	89.72
Supervised	91.24	89.22	79.23	92.14	82.41	90.83

Qualitative Evaluations

Citation

  
    @article{zhang2023improving,
        title={Improving the Generalization of Segmentation Foundation Model under Distribution Shift via Weakly Supervised Adaptation},
        author={Zhang, Haojie and Su, Yongyi and Xu, Xun and Jia, Kui},
        journal={arXiv preprint arXiv:2312.03502},
        year={2023}
    }