Research map / Detection and segmentation / Weakly supervised detection and segmentation
Test Time Adaptation With Regularized Loss for Weakly Supervised Salient Object Detection
Olga Veksler
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| adapt the base CNN on augmented copies of each test image using regularized loss and per-image hyperparameters | weakly supervised SOD model [33] · DUTS, ECSSD, DUTO, PascalS, HKUIS, THUR | Fβ: [33] -> ours+h: .753 -> .795 (DUTS); .868 -> .911 (ECSSD); .697 -> .752 (DUTO); .796 -> .807 (PascalS); .866 -> .899 (HK | Improved |
| set λm adaptively per test image instead of using a fixed value (ablation) | test-time adaptation on DUTO · DUTO (500-image sample) | Fβ: fixed settings .741, .738, .752, .755, .742, .732, .728, .718, .703; adaptive .765 | Improved |
| remove border loss (ablation) | test-time adaptation · DUTS, ECSSD, DUTO, PascalS, HKUIS, THUR | Fβ: DUTS .795 -> .789; ECSSD .911 -> .910; DUTO .752 -> .748; PascalS .807 -> .800; HKUIS .899 -> .892; THUR .760 -> .755 | Hurt |
| vary sparse-CRF σ from 0.15 to 0.1 or 0.20 (ablation) | test-time adaptation · DUTS, ECSSD, DUTO, PascalS, HKUIS, THUR | Fβ: DUTS .795 -> .789/.797; ECSSD .911 -> .894/.915; DUTO .752 -> .751/.743; PascalS .807 -> .803/.803; HKUIS .899 -> .902/. | Mixed |
| increase test-time dataset size from 16 to 32 images (ablation) | test-time adaptation · ECSSD | Fβ: .910 -> .915 | Improved |
| increase test-time dataset size from 16 to 64 images (ablation) | test-time adaptation · ECSSD | Fβ: .910 -> .913 | No clear effect |
| reduce adaptation to 2 or 5 epochs (ablation) | test-time adaptation · ECSSD | Fβ: 10 epochs .910; 2 epochs .873; 5 epochs .883 | Hurt |
| increase adaptation to 20 or 30 epochs (ablation) | test-time adaptation · ECSSD | Fβ: 10 epochs .910; 20 epochs .915; 30 epochs .915 | Improved |
| increase adaptation to 50 epochs (ablation) | test-time adaptation · ECSSD | Fβ: 10 epochs .910 -> 50 epochs .912 | No clear effect |
| use Gaussian noise augmentation instead of small shifts (ablation) | test-time adaptation · ECSSD | Fβ: .911 -> .854 | Hurt |
| use color perturbation instead of small shifts (ablation) | test-time adaptation · ECSSD | Fβ: .911 -> .897 | Hurt |
| add color perturbation to small-shift augmentation (ablation) | test-time adaptation · ECSSD | Fβ: .911 -> .909 | No clear effect |
Closest papers by meaning
- Test-Time Adaptive Object Detection with Foundation Model (NeurIPS 2025 · similarity 0.88)
- Structure-Consistent Weakly Supervised Salient Object Detection with Local Saliency Coherence (AAAI 2021 · similarity 0.86)
- Weakly-Supervised Salient Object Detection Using Point Supervision (AAAI 2022 · similarity 0.86)
- Training-Time-Friendly Network for Real-Time Object Detection (AAAI 2020 · similarity 0.85)
- Self-Bootstrapping for Versatile Test-Time Adaptation (ICML 2025 · similarity 0.85)
- What How and When Should Object Detectors Update in Continually Changing Test Domains? (CVPR 2024 · 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.
