Research map / Detection and segmentation / Weakly supervised detection and segmentation
Unsupervised Salient Instance Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SCoCo, an unsupervised salient instance detection method that models saliency coherence and contrast in frozen DINO ViT features using a global background adaptation module and a locality-aware similarity module with contrastive losses.
Problem
Unsupervised salient instance detection: existing unsupervised salient object detection only yields region-level saliency, without distinguishing individual salient instances, and salient instance detection methods require pixel-, box-, or image-level supervision.
Approach
Exploits saliency coherence within regions and saliency contrast across regions in self-supervised transformer features; a GBA module learns adaptive saliency thresholds with a scene-level contrastive loss, and a LAS module learns instance grouping with an instance-level contrastive loss, using pseudo labels from Normalized Cut on DINO features.
Key results
- SCoCo achieves 70.7% mAP50 on ILSO-1K, 64.8% on ILSO-2K, and 71.4% on SIS, outperforming unsupervised baselines such as FOUND (67.1% on SIS mAP50).
- SCoCo surpasses all weakly-supervised methods (e.g., WSID-Net 68.4% mAP50 on SIS) and is comparable to fully-supervised methods (OQTR 88.1% on SIS mAP50).
- Ablation on SIS shows the full model reaches 71.4% mAP50 versus 64.7% for the baseline without GBA and LAS modules.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SCoCo instead of the strongest weakly supervised baseline, WSID-Net | salient instance detection · ILSO-1K test | mAP50: 68.3% -> 70.7% | Helped |
| use SCoCo instead of the strongest weakly supervised baseline, WSID-Net | salient instance detection · ILSO-2K test | mAP50: 59.4% -> 64.8% | Helped |
| use SCoCo instead of the strongest weakly supervised baseline, WSID-Net | salient instance detection · SIS test | mAP50: 68.4% -> 71.4% | Helped |
| add GBA (ablation) | SCoCo · SIS test | mAP50: 64.7% -> 67.0% | Helped |
| add the scene-level contrastive loss to GBA (ablation) | SCoCo with GBA · SIS test | mAP50: 67.0% -> 68.4% | Helped |
| add LAS (ablation) | SCoCo · SIS test | mAP50: 64.7% -> 67.4% | Helped |
| add the instance-level contrastive loss to LAS (ablation) | SCoCo with LAS · SIS test | mAP50: 67.4% -> 69.0% | Helped |
| combine GBA, scene-level loss, LAS, and instance-level loss (ablation) | SCoCo · SIS test | mAP50: 64.7% -> 71.4% | Helped |
| initialize the ViT-S backbone with DINO rather than MoCo-v3 (ablation) | SCoCo · SIS test | mAP50: 70.5% -> 71.4% | Helped |
| initialize the ViT-S backbone with DINO rather than SWaV (ablation) | SCoCo · SIS test | mAP50: 70.2% -> 71.4% | Helped |
| use ViT-B instead of ViT-S with DINO initialization (ablation) | SCoCo · SIS test | mAP50: 71.4% -> 71.9% | Helped |
| use ViT-B instead of ViT-S with DINO initialization (ablation) | SCoCo · SIS test | mAP70: 55.7% -> 55.6% | No clear effect |
Surprising findings
- Fully-supervised methods are not uniformly better: on ILSO-1K/SIS, SCoCo (unsupervised) is comparable to some fully-supervised methods, while OQTR remains stronger on SIS (88.1% mAP50).
Limitations
- May fail to detect small salient instances because it relies on high-level, low-resolution features.
- Evaluated only on three SID datasets with mAP metrics.
Experimental setup
Datasets: ILSO-1K ILSO-2K SIS
Benchmarks: ILSO-1K ILSO-2K SIS
Metrics: mAP50 mAP70
Baselines: FreeSOLO CutLER VitDet SelfMask FOUND MSRNet S4Net SCG MDNN RDPNet OQTR MAP
Builds on: DINO Mask-RCNN CutLER FOUND
Closest papers by meaning
- Instance-Level Salient Object Segmentation (CVPR 2017 · similarity 0.90)
- Deep Unsupervised Saliency Detection: A Multiple Noisy Labeling Perspective (CVPR 2018 · similarity 0.89)
- Structure-Consistent Weakly Supervised Salient Object Detection with Local Saliency Coherence (AAAI 2021 · similarity 0.88)
- Texture-Guided Saliency Distilling for Unsupervised Salient Object Detection (CVPR 2023 · similarity 0.88)
- DeepUSPS: Deep Robust Unsupervised Saliency Prediction via Self-supervision (NeurIPS 2019 · similarity 0.88)
- Unsupervised Domain Adaptive Salient Object Detection through Uncertainty-Aware Pseudo-Label Learning (AAAI 2022 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.