Research map / Detection and segmentation / CNN features
FOCUS: Towards Universal Foreground Segmentation
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 FOCUS, a unified foreground segmentation framework (COD, SOD, SD, DBD, FD) that uses DINOv2 features with edge enhancement, ground queries for foreground and background, and a frozen-CLIP contrastive refiner.
Problem
Foreground segmentation tasks (SOD, COD, shadow, defocus blur, forgery detection) are typically handled by task-specific architectures, and prior methods focus on recognizing foreground without explicitly modeling the background or its relation to the foreground.
Approach
A Mask2Former-style transformer decoder with two learned "ground queries" (foreground and background) initialized from PCA-binarized DINOv2 features as attention masks, an edge enhancer that fuses ResNet50 edge features via ViT-Adapter-style injector/extractor, and a CLIP refiner that contrastively aligns mask-fused images with foreground/background text prompts to sharpen boundaries.
Key results
- On COD, FOCUS reports S-measure 0.912 on CAMO and 0.910 on COD10K, outperforming BiRefNet on most metrics.
- On forgery detection (CASIA-1.0), FOCUS reaches F1 0.892 versus 0.654 for the prior best EVPv2, a 23.8% gain as reported.
- On ISTD shadow detection, FOCUS achieves BER 0.98 versus 1.11 for SILT (the previous best).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add joint foreground-background prediction (ablation) | FOCUS with DINOv2-L · CAMO and PASCAL-S | S-measure: .853 -> .854; .851 -> .853 | Helped |
| add CLIP refiner (ablation) | FOCUS with DINOv2-L and joint prediction · CAMO and PASCAL-S | S-measure: .854 -> .861; .853 -> .855 | Helped |
| add edge enhancer (ablation) | FOCUS with DINOv2-L and CLIP refiner · CAMO and PASCAL-S | S-measure: .861 -> .872; .855 -> .870 | Helped |
| use ADE20K-pretrained DINOv2-G (ablation) | FOCUS · CAMO and PASCAL-S | S-measure: .897 -> .901; .889 -> .896 | Helped |
| freeze the DINOv2-G backbone (ablation) | FOCUS · CAMO | S-measure: .909 -> .905 | Hurt |
| freeze the DINOv2-G backbone (ablation) | FOCUS · PASCAL-S | S-measure: .897 -> .893 | Hurt |
| use FOCUS instead of the strongest listed task-specific baseline | FOCUS · CAMO, NC4K, DUTS-TE, DUT-OMRON, ECSSD, PASCAL-S, ISTD, DUT, CUHK, and CASIA-1.0 | primary reported metric: .904 -> .912; .914 -> .915; .921 -> .929; .862 -> .868; .941 -> .943; .879 -> .898; 1.11 -> 0.98; .891 -> .912; .932 -> | Helped |
Surprising findings
- Initializing the decoder with PCA-reduced, binarized DINOv2 features lets the feature maps approach ground-truth quality even without further forward propagation, per the authors.
- Freezing the DINOv2-G backbone (0.1G trainable parameters) yields only a slight drop versus full fine-tuning and still matches or beats BiRefNet and SelfReformer on several metrics.
Limitations
- Background segmentation is treated as a separate task, and the CLIP refiner's prompts must be manually adjusted per task.
- Some results are reported only on a subset of tasks and comparisons with unified baselines are limited to EVPv1/EVPv2.
- Some reported metrics are fine-tuned with a DINOv2-G backbone pre-trained on ADE20K, so gains may partly reflect backbone scale.
Experimental setup
Datasets: CAMO-TR COD10K-TR DUTS-TR ISTD CUHK DUT CASIA-2.0 ADE20K
Benchmarks: CAMO COD10K CHAMELEON NC4K DUTS-TE DUT-OMRON HKU-IS ECSSD PACAL-S ISTD DUT CUHK
Metrics: S-measure E-measure weighted F-measure MAE BER F-measure F1 AUC
Baselines: SINet PFNet ZoomNet BSA-Net FSPNet ZoomNeXt BiRefNet EVPv1 EVPv2 SelfReformer MENet VST
Builds on: Mask2Former DINOv2 CLIP ViT-Adapter DETR
Closest papers by meaning
- Visual Boundary Knowledge Translation for Foreground Segmentation (AAAI 2021 · similarity 0.89)
- Attention-Guided Unified Network for Panoptic Segmentation (CVPR 2019 · similarity 0.89)
- Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling (CVPR 2024 · similarity 0.86)
- Bridge the Points: Graph-based Few-shot Segment Anything Semantically (NeurIPS 2024 · similarity 0.85)
- Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · similarity 0.85)
- UNIP: Rethinking Pre-trained Attention Patterns for Infrared Semantic Segmentation (ICLR 2025 · similarity 0.85)
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.