We present SCoRe (Support-Calibrated Role fusion), a framework to utilize two complementary vision foundation model-based (VFM) segmenters for training-free cross-domain few-shot segmentation (CD-FSS). Our approach is based on the observation that current VFMs for CD-FSS possess capabilities where one model locates the correct object, while another model computes the better contour. Our approach uses: 1. a method to identify a VFM segmenter to lead Few-Shot Segmentation (FSS) task for a given concept, 2. a novel deterministic boundary correction method using bands of near- and far-pixels at the lead segmenter's predicted boundary of an object, and 3. a region correction mechanism for the other segmenter to conservatively prune components, fill holes, and bridge short gaps on the lead segmenter's predicted masks to improve accuracy. We evaluate SCoRe using RSRM and FSS-SAM3 across 3,314 matched five-shot episodes from 10 datasets. SCoRe achieves 72.738% foreground mIoU averaged across datasets, a gain of 5.210 percentage points over RSRM (paired 95% CI [4.618, 5.972]), with per-dataset gains up to 8.461 points over the strongest standalone method and the best point estimate on 7 of 10 datasets.
Figures & tables
Fig. 1: Overview of SCoRe and the complementary strengths of its experts. (A) Five leave-one-support-out folds compare the two completed routes; a majority vote selects and freezes the route for query inference. (B) RSRM more accurately identifies the target region but produces imprecise contours, whereas FSS-SAM3 provides more accurate boundaries but includes false-positive regions (the extra building in orange). The selected region-led route preserves RSRM’s semantic extent and uses FSS-SAM3 to refine its contours, producing a mask with substantially higher region overlap and boundary accuracy in the illustrated example. Support labels are used only for route calibration; query ground truth is used only for visualization and evaluation.
Fig. 2: Qualitative results showing how SCoRe combines the two experts. In (a) R-LED, RSRM identifies the target region and FSS-SAM3 refines its boundaries. In (b) B-LED, FSS-SAM3 provides the contours and RSRM helps remove false regions and recover missing foreground. Green marks correct foreground, orange false positives, and blue missed foreground. Scores below the masks show IoU (%); the green boxes above the SCoRe predictions show gains over the best displayed baseline in percentage points.
Dataset
Episodes
R/B
RSRM
FSS-SAM3
HERA
Matcher
GF-SAM
PerSAM-5S
SCoRe
Δ [95% CI]
Chest X-ray
60
1/0
89.104
86.107
88.671
79.990
37.207
31.258
90.092
+0.988[+0.311,+1.621]
KSDD2
60
1/0
64.386
51.844
65.867
13.339
27.966
5.766
74.327
+8.461[+2.840,+15.016]
ISIC 2018
60
3/0
68.539
67.433
66.863
38.719
43.516
17.761
73.374
+4.835[+3.081,+7.169]
Oxford Pet
74
6/31
76.051
82.744
75.592
77.217
79.572
67.277
82.702
−0.042[−0.102,−0.00029]
DeepGlobe
60
6/0
57.981
48.242
56.724
53.429
53.438
41.897
58.413
+0.432[−1.066,+1.908]
Kvasir
600
2/1
72.267
71.102
68.750
35.722
29.731
18.950
72.870
+0.603[−1.264,+2.483]
TABLE I: Matched five-shot foreground mIoU (%). Bold and underline denote the best and second-best method. R/B counts region-led and boundary-led routing units. Δ is the difference against the strongest standalone method in each row, with a paired 95% CI, in percentage points.
Dataset
RSRM
FSS-SAM3
HERA
FROST
SCoRe
DeepGlobe
48.32/59.04
30.93/46.17
43.18/59.62
50.53 / 59.84
48.12/58.15
ISIC 2018
57.62/65.67
64.41 /70.21
52.42/61.78
55.50/66.68
61.98/ 71.36
Chest X-ray
87.57/88.92
87.02/82.58
75.70/87.46
84.69/86.06
90.15 / 88.98
FSS-1000
80.68/83.06
88.06/91.24
80.15/82.42
82.92/86.26
88.87 / 91.36
Mean
68.55/74.17
67.60/72.55
62.86/72.82
68.41/74.71
72.28 / 77.46
TABLE II: Single-seed mIoU (%), shown as 1-shot/5-shot. Each setting uses 1,200 episodes per dataset (2,400 for FSS-1000). The mean weights datasets equally. Bold marks the best score per shot.
Vision foundation models (VFMs) have achieved strong performance across various vision tasks. However, it still remains challenging to apply VFMs for cross-domain few-shot segmentation (CD-FSS), which segments objects of novel classes under domain shifts using only a few labeled exemplars. The challenge is mainly driven by two factors: (1) limited labeled exemplars per novel class relative to the scale of VFM pre-training, making the model prone to overfitting during retraining, and (2) target-domain shifts underrepresented during pre-training, inducing cross-domain inconsistency and layer-wise sensitivity. To address these issues, we propose Hierarchical Exemplar Representation Adaptation (HERA), a three-stage select-regularize-calibrate VFM-based segmentation framework that learns effectively from limited labels and adapts to novel domains without source-data retraining. We first design Hierarchical Layer Selection (HLS) to adaptively identify the most informative VFM layer using a data-dependent Exemplar Transfer Risk (ETR) computed for each candidate layer. Then, Prior-Guided Regularization (PGR) regularizes interactions on the selected representation, yielding well-structured local signals for the subsequent stage. Furthermore, Pixelwise Adaptive Calibration (PAC) combines the selected representation with the refined interaction maps to calibrate pixel-wise predictions, producing consistent masks. Together, these stages form a hierarchical select-regularize-calibrate pipeline that guides frozen VFM features in new domains while fine-tuning less than 2.7% of parameters at test time. Extensive experiments show that HERA surpasses the state of the art by more than 4.1 mIoU across multiple CD-FSS benchmarks.
Cross-domain Few-shot Segmentation (CD-FSS) aims to transfer knowledge learned from source domain to distinct target domains, segmenting unseen target classes with only a few annotated samples. Although existing methods have made significant progress, they still rely on training or fine-tuning processes, which incur high computational costs and risk overfitting. We observe that when powerful and general-purpose vision foundation models are incorporated into these methods, their performance shows only marginal improvement or even degrades due to overfitting. To address this, we eliminate trainable parameters and propose a training-free framework to avoid both training overhead and overfitting. Built upon the self-supervised vision encoder DINOv3, our framework addresses cross-domain challenges through three core modules. First, the Semantic-aware Feature Re-fusion (SAFR) module identifies and re-fuses features that emphasize semantic patterns, generating representations with enhanced semantic discriminability. Additionally, the Adaptive Support Enhancement (ASE) module narrows semantic gaps between support and query through robust query information aggregation. Finally, the Hybrid Prototype Matching (HPM) module integrates matching results from diverse prototypes to adapt to varying semantic complexity across domains. Extensive experiments on four target domain datasets demonstrate that our method achieves state-of-the-art performance in CD-FSS without any training.
Sujun Sun, Mingwu Ren, Haofeng Zhang
School of Computer Science and Engineering, Nanjing University of Science and Technology, China · State Key Laboratory of Intelligent Manufacturing of Advanced Construction Machinery, China
Cross-domain Few-shot Segmentation (CD-FSS) aims to learn generalizable segmentation capability from abundant annotated samples in the source domain, enabling accurate segmentation of novel classes in the target domain with only a few annotated samples. Existing CD-FSS methods mainly focus on mitigating feature distribution shifts caused by style gaps while ignoring significant differences in class semantic granularity and discriminative attributes across domains, leading to two key degradations in support-query matching: semantic over-alignment and attribute over-alignment. To this end, we propose the Dual Hierarchical Aggregation Network (DHANet), which comprises three key modules. First, the Hierarchical Spatial Aggregation (HSA) module performs multi-scale region aggregation of pixel features along the spatial dimension, generating hierarchical semantic-enhanced features to alleviate semantic over-alignment. Additionally, the HCA module conducts multi-scale attribute aggregation along the channel dimension, generating hierarchical attribute-enhanced features to mitigate attribute over-alignment. Finally, we propose the Online Probabilistic Semantic Bank (OPSB), which progressively constructs and updates class probability distributions from query predictions during inference, and samples multiple pseudo-prototypes as additional support information to mitigate insufficient support. Extensive experiments on four target-domain datasets demonstrate that our method achieves state-of-the-art performance.
Sujun Sun, Mingwu Ren, Haofeng Zhang
School of Computer Science and Engineering, Nanjing University of Science and Technology, China · State Key Laboratory of Intelligent Manufacturing of Advanced Construction Machinery, China