Research map / Detection and segmentation / Open-set and few-shot detection
Unsupervised Image Matching and Object Discovery as Optimization
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use normalized rather than unnormalized stand-out scores (ablation) | Proposed method · ObjectDiscovery and VOC_6x2, separate | CorLoc: 81.9 $\pm$ 0.9 -> 83.1 $\pm$ 0.8; 65.9 $\pm$ 1.0 -> 67.2 $\pm$ 1.0 | Helped |
| Add ensemble postprocessing (ablation) | Proposed method · ObjectDiscovery and VOC_6x2, separate | CorLoc: 82.9 $\pm$ 0.8 -> 83.8 $\pm$ 0.2; 66.6 $\pm$ 0.7 -> 67.4 $\pm$ 0.4 | Helped |
| Initialize greedy ascent with continuous optimization (ablation) | Proposed method · VOC_all, separate | CorLoc: 36.4 $\pm$ 0.3 -> 37.8 $\pm$ 0.3 | Helped |
| Add ensemble postprocessing (ablation) | Proposed method · VOC_all, separate | CorLoc: 37.8 $\pm$ 0.3 -> 39.2 $\pm$ 0.2 | Helped |
| Add continuous optimization (ablation) | Proposed method · ObjectDiscovery, VOC_6x2, and VOC_all, mixed | CorLoc: 83.0 $\pm$ 0.4 -> 80.8 $\pm$ 0.5; 60.2 $\pm$ 0.4 -> 59.3 $\pm$ 0.4; 39.8 $\pm$ 0.2 -> 38.5 $\pm$ 0.2 | Hurt |
| Increase the candidate-region limit nu from 1 to 10 (ablation) | Proposed method · VOC_6x2, separate | CorLoc: 68.1 $\pm$ 0.7 -> 70.0 $\pm$ 0.3 | Helped |
| Replace WHO descriptors with VGG16 conv5_3 ROI-pooled, unnormalized features (ablation) | Proposed method · VOC_6x2, separate | CorLoc: 68.8 $\pm$ 0.5 -> 70.7 $\pm$ 0.2 | Helped |
| Replace WHO descriptors with the other tested deep-feature configurations (ablation) | Proposed method · VOC_6x2, separate | CorLoc: 68.8 $\pm$ 0.5 vs. 64.2 $\pm$ 0.2; 57.1 $\pm$ 0.6; 63.1 $\pm$ 0.2; 63.4 $\pm$ 0.4; 64.9 $\pm$ 0.2; 64.1 $\pm$ 0.4; 68.2 | Hurt |
| Use selective search quality proposals instead of randomized Prim's proposals (ablation) | Proposed method · VOC_6x2, separate | CorLoc: 69.4 $\pm$ 0.4 -> 62.8 $\pm$ 0.6 | Hurt |
| Use selective search quality proposals rather than randomized Prim's proposals (ablation) | Proposed method · VOC_6x2, separate | CorLoc: Cho et al.: 67.6 -> 32.6; ours: 69.4 $\pm$ 0.4 -> 62.8 $\pm$ 0.6 | Mixed |
| Use the full optimization-based method with normalized stand-out, CO, and EM rather than Cho et al. | Proposed method · ObjectDiscovery and VOC_6x2, separate | CorLoc: 84.2 -> 85.8 $\pm$ 0.6; 67.7 -> 69.4 $\pm$ 0.3 | Helped |
| Use the proposed method without continuous optimization rather than Cho et al.'s execution | Proposed method · ObjectDiscovery, VOC_6x2, and VOC_all, mixed | CorLoc: 82.2 -> 83.0 $\pm$ 0.4; 55.9 -> 60.2 $\pm$ 0.4; 37.5 -> 39.8 $\pm$ 0.2 | Helped |
Closest papers by meaning
- Large-Scale Unsupervised Object Discovery (NeurIPS 2021 · similarity 0.90)
- Unsupervised Learning of Object Landmarks via Self-Training Correspondence (NeurIPS 2020 · similarity 0.89)
- unMORE: Unsupervised Multi-Object Segmentation via Center-Boundary Reasoning (ICML 2025 · similarity 0.86)
- Unsupervised Object Detection with Theoretical Guarantees (NeurIPS 2024 · similarity 0.86)
- OmniGlue: Generalizable Feature Matching with Foundation Model Guidance (CVPR 2024 · similarity 0.86)
- Pose-Guided Self-Training with Two-Stage Clustering for Unsupervised Landmark Discovery (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.