Co-salient Object Detection (CoSOD) aims to segment salient objects that consistently appear across a group of related images. Despite the notable progress achieved by recent training-based approaches, they still remain constrained by the closed-set datasets and exhibit limited generalization. However, few studies explore the potential of Vision Foundation Models (VFMs) to address CoSOD, which demonstrate a strong generalized ability and robust saliency understanding. In this paper, we investigate and leverage VFMs for CoSOD, and further propose a novel training-free method, TF-SSD, through the synergy between SAM and DINO. Specifically, we first utilize SAM to generate comprehensive raw proposals, which serve as a candidate mask pool. Then, we introduce a quality mask generator to filter out redundant masks, thereby acquiring a refined mask set. Since this generator is built upon SAM, it inherently lacks semantic understanding of saliency. To this end, we adopt an intra-image saliency filter that employs DINO's attention maps to identify visually salient masks within individual images. Moreover, to extend saliency understanding across group images, we propose an inter-image prototype selector, which computes similarity scores among cross-image prototypes to select masks with the highest score. These selected masks serve as final predictions for CoSOD. Extensive experiments show that our TF-SSD outperforms existing methods (e.g., 13.7% gains over the recent training-free method). Codes are available at https://github.com/hzz-yy/TF-SSD.
Figures & tables
Figure 1: Segmentation results derived from SAM. The oracle masks selected under the ground truth guidance exhibit co-salient representations, which are highlighted in the green rectangle.
Figure 2: Comparison of our TF-SSD and other CoSOD methods. Fm and Sm denote the F-measure and S-measure, respectively. Our method surpasses training-based methods and achieves 13.7% improvement over the recent training-free SOTA method.
Figure 3: Overview pipeline of our TF-SSD. It contains four components: a Quality Mask Generator (QMG) that filters out exhaustive candidate masks from SAM, a DINO encoder for salient attention and semantic prototype extraction, an Intra-image Saliency Filter (ISF) for visually salient purification within a single image, and an Inter-image Prototype Selector (IPS) that builds the saliency relationship across group images to discover the co-salient masks as the final predictions for CoSOD. For clarity, our notations are defined as: mn,t denotes an individual mask, Mn denotes the mask set of image n (e.g., Mnraw , Mnrefine ).
Algorithm 1 Quality Mask Generator (QMG)
Figure 4: Illustration of Intra-image Saliency Filter. Row 1 : Original image, self-attention map of DINO, and ground truth. Row 2 : Three mask proposals generated by QMG. Row 3 : Attention response obtained by element-wise multiplication of each mask with the attention map. Mask 1 achieves a high saliency score due to strong overlap between salient regions and the mask proposal, while Masks 2 and 3 receive low scores due to non-overlap.
Figure 5: Illustration of Inter-image Prototype Selector (IPS) with an example of N=3 images and T=3 masks per image. A pairwise similarity matrix is constructed using mask prototypes, where different colors denote prototypes from different images. For each candidate (e.g., p1,1 of image I1 ), we select its maximum score (deep red cells) against the candidates from each of the other images ( I2 and I3 ). These maximum scores are summed as the total co-saliency score smaxco . The mask with the highest smaxco is selected as the final prediction.
Component
CoCA [ 44 ]
CoSal2015 [ 41 ]
CoSOD3k [ 8 ]
ID
QMG-1
QMG-2
QMG-3
ISF
IPS
MAE↓
Fβmax↑
Eξmax↑
Sα↑
MAE↓
Fβmax↑
Eξmax↑
Sα↑
MAE↓
Fβmax↑
Eξmax↑
Sα↑
1
✓
✓
0.171
0.447
0.662
0.578
0.133
0.680
0.779
0.704
0.139
0.610
0.752
0.667
2
✓
✓
✓
0.121
0.497
0.747
0.650
0.095
0.815
0.875
0.799
0.099
0.744
0.839
0.754
3
✓
✓
✓
✓
0.119
0.522
0.765
0.634
0.092
0.850
0.900
0.818
0.099
0.785
0.862
0.786
5
✓
✓
✓
✓
✓
0.077
0.686
0.815
0.763
0.089
0.899
0.926
0.890
0.090
0.860
0.908
0.861
Table 2: Ablation study of the proposed components in TF-SSD on three benchmark datasets. QMG-1, QMG-2, and QMG-3 denote the three stages of our QMG, respectively. The final row represents our full framework. Best results are marked bold .
Figure 6: Qualitative comparison. We compare the result visualization between our method and previous SOTA methods. Among them, ZS-CoSOD [ 34 ] is the training-free method. GCoNet+ [ 45 ] , VCP [ 31 ] and SCoSPARC [ 3 ] are training-based methods. It’s observed that TF-SSD obtains precise co-salient masks compared to other SOTA methods, even in complicated cases ( e.g . , “Chopsticks” and “Axe”).
Strategy
CoCA
CoSal2015
CoSOD3k
Fβmax↑
Eξmax↑
Fβmax↑
Eξmax↑
Fβmax↑
Eξmax↑
Avg
0.568
0.778
0.775
0.826
0.737
0.792
Max
0.686
0.815
0.899
0.926
0.860
0.908
Table 3: Ablation study of the N−1 similarity score selection (Eq. 9 ) in IPS. We compare using the maximum similarity ( Max ) versus the average similarity ( Avg ) to compute the similarity score among cross-image prototypes.
T
MAE ↓
Fβ max↑
Eξmax↑
Sα↑
10
0.078
0.616
0.786
0.718
8
0.078
0.645
0.789
0.726
7
0.077
0.666
0.809
0.744
6
0.077
0.686
0.815
0.763
5
0.077
0.667
0.804
0.748
Table 4: Ablation study on the number of selected masks ( T ) in the ISF module . Performance is evaluated on the CoCA dataset.
τarea
Fβmax
Oracle Fβmax
0.002
0.545
0.696
0.005
0.601
0.722
0.01
0.686
0.768
0.02
0.653
0.737
Table 5: Ablation study on different area ratio thresholds τarea on the CoCA dataset.
Backbone
Fβmax
Eξmax
Sα
DINO ViT-B/16
0.556
0.742
0.697
DINOv2 ViT-B/14
0.607
0.783
0.708
DINO ViT-B/8
0.686
0.815
0.763
Table 6: Performance comparison of different DINO backbones on the CoCA dataset.
α
β
Fβmax
Eξmax
Sα
0.0
1.0
0.473
0.702
0.611
0.3
0.7
0.591
0.737
0.651
0.5
0.5
0.658
0.795
0.744
0.7
0.3
0.686
0.815
0.763
1.0
0.0
0.641
0.782
0.735
Table 7: Ablation on quality score weights on the CoCA dataset.
rmin
rmax
Fβmax
Eξmax
Sα
0.05
0.85
0.661
0.798
0.747
0.1
0.6
0.678
0.808
0.756
0.15
0.7
0.686
0.815
0.763
0.2
0.5
0.619
0.741
0.669
Table 8: Ablation on ideal size range on the CoCA dataset.
τcon
Fβmax
Eξmax
Sα
0.5
0.652
0.791
0.741
0.7
0.673
0.805
0.754
0.85
0.686
0.815
0.763
0.95
0.679
0.796
0.756
Table 9: Ablation on overlap threshold on the CoCA dataset.
Figure 7: Visualization of SAM mask proposals for ”Accordion” and ”bicycle” image groups. For each group, we show the original images (row 1), ground truth(GT) masks (row 2), and the top-5 mask proposals ranked by quality score Sn,tba (rows 3-7). Green borders indicate the final predictions selected by our TF-SSD pipeline.
Figure 8: Visualization of DINO attention maps for three categories: ”airplane”, ”butterfly”, and ”penguin”. Row 1: Resized input images. Row 2: Resized GT masks. Row 3: DINO attention maps An that naturally highlight salient objects.
Figure 9: Limitation on small object detection. Visualizations of failed cases from the ”moon” category.
Co-salient object detection (Co-SOD) requires a model to find foreground regions that are salient in individual images and supported by the image group. We present Rank-Consistent Set Reasoning (RCSR), a supervised dense-prediction framework that models a group as an unordered set rather than as a sequence of images or a semantic label. The core idea is to rank how strongly each spatial region agrees with a small collection of learned group slots at every image scale, and to aggregate these ranks with a robust trimmed statistic. This suppresses accidental pairwise matches and prevents one atypical group member from dominating the shared representation. A set encoder builds group slots directly from multi-scale visual features, while a rank-consistency gate measures whether the ordering of candidate regions is stable across group members. The gated slots are decoded jointly with per-image features to produce co-saliency maps. The model contains no natural-language branch, no open-vocabulary detector, and no external segmentation model. We further introduce a group permutation objective and hard-distractor augmentation so that the model learns the properties of a set-level target rather than memorizing image order or isolated visual saliency. We formulate an evaluation protocol for CoCA, CoSal2015, and CoSOD3k, together with tests of group-size robustness, distractor rejection, order invariance, and cross-dataset transfer.
Yuan Xiang, Matteo Rossi, Yingzhou Chen
University of California, Los Angeles (UCLA) USA · Polytechnic University of Turin Italy
Salient Object Detection (SOD) remains an essential yet underexplored task in the era of large-scale vision models. Although foundation models like SAM exhibit strong generalization, their potential for SOD is not fully realized, and training or fully fine-tuning them is computationally expensive and prone to overfitting under limited data. To overcome these challenges, we introduce GLASSNet, a Global-Local feature decoding framework that uses SAMv2 as a frozen encoder paired with a lightweight, spatially aware convolutional adapter-reducing learnable encoder parameters by over 97%. To enhance saliency quality, GLASSNet employs a dual-decoder architecture: one decoder captures global, long-range semantics with an expanded receptive field, while the other captures fine local details such as edges and textures. Fusing these complementary cues yields saliency maps that combine global coherence with local precision, producing accurate final masks. Extensive experiments on standard SOD and camouflaged object detection benchmarks show that GLASSNet surpasses state-of-the-art methods, demonstrating the power of frozen foundation models combined with targeted adaptation and global-local decoding.
Morteza Moradi, Mohammad Moradi, Simone Palazzo +2
University of Catania, Catania, Italy · Quintic AI, San Francisco, CA, USA
Referring Camouflaged Object Detection (Ref-COD) requires segmenting hidden targets guided by reference cues. While supervised methods are annotation-heavy and training-free approaches via sparse point-prompting are sensitive to localization errors, we propose GFR-SAM, a robust three-stage training-free framework. GFR-SAM shifts the paradigm from fragile point-matching to a "Generate-Filter-Refine" pipeline. First, we introduce In-Context Exemplar-guided Segmentation, empowering SAM3 with cross-image inference to generate candidate masks via holistic visual exemplars, bypassing its native intra-image constraints. Second, a Region-Global Contrastive Filtering module ranks candidates through DINOv3-based prototypical alignment, effectively suppressing background distractors. Finally, a Geometric-Semantic Refinement module synergizes bounding box and text prompts to recover fine-grained boundaries and enhance instance recall. Evaluated on the R2C7K benchmark, GFR-SAM outperforms existing training-free methods by 8.7% in weighted F-measure (Fβw) and competes with supervised state-of-the-art counterparts. Ultimately, this work underscores the potential of unlocking SAM3's latent capability for cross-image In-Context prompting, establishing a robust, training-free paradigm that effectively bridges the gap between general-purpose foundation models and specialized, label-intensive perception tasks without the need for task-specific fine-tuning.
Yilong Yang, Jianxin Tian, Shengchuan Zhang +1
Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University