Research map / Detection and segmentation / Open-set and few-shot detection
Evaluating Object-Centric Models beyond Object Discovery
Krishnakant Singh, Simone Schaub-Meyer, Stefan Roth
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| add DINOv2 feature reconstruction decoder (ablation) | StableLSD · eGQA (Phi2) | mIoU: 47.94 -> 55.76 | Improved |
| add HOG feature decoder on top of image+feature decoders (ablation) | mFRESA · eGQA (Phi2) | AwGA: 13.37 -> 13.91 | Improved |
| add feature decoder (no HOG) (ablation) | StableLSD · VQAv2 (Phi2) | accuracy: 62.06 -> 61.40 | No clear effect |
| add both feature and HOG decoders (ablation) | StableLSD · VQAv2 (Phi2) | accuracy: 62.06 -> 65.58 | Improved |
| add both feature and HOG decoders (ablation) | StableLSD · GQA (Phi2) | accuracy: 51.45 -> 53.90 | Improved |
| use feature reconstruction instead of image reconstruction | OCL models · robustness benchmarks | accuracy | Improved |
| use finetuned DINOv2 (FT-DINOSAUR) instead of frozen | DINOSAURv2 · VQAv2 (Phi2) | accuracy: 63.84 -> 60.97 | Hurt |
| use integrated-gradient instead of gradient attribution (ablation) | AwGA · eGQA | model ranking: Spearman 0.77 | No clear effect |
| use Qwen2-7B instead of Phi2-3B as LLM (ablation) | AwGA · eGQA | rank correlation: Spearman 0.85 | No clear effect |
| use 1-layer MLP instead of 2-layer MLP connector (ablation) | AwGA · eGQA | rank correlation: Spearman 0.74 | No clear effect |
| use OCL slots (7 tokens) instead of DINOv2 (196 tokens) | VLM · VQAv2 (Phi2) | accuracy: 71.15 -> 65.58 | Hurt |
| replace vision tokens by noise (Blind-VLM) | VLM · Boolean CVQA (Phi2) | accuracy: 63.72 vs 71.41 (blind better) | Improved |
Closest papers by meaning
- Bootstrapping Top-down Information for Self-modulating Slot Attention (NeurIPS 2024 · similarity 0.89)
- Multi-Part Object Representations via Graph Structures and Co-Part Discovery (NeurIPS 2026 · similarity 0.88)
- Learning to Compose: Improving Object Centric Learning by Injecting Compositionality (ICLR 2024 · similarity 0.88)
- Exploring the Effectiveness of Object-Centric Representations in Visual Question Answering: Comparative Insights with Foundation Models (ICLR 2025 · similarity 0.87)
- Object-Centric Pretraining via Target Encoder Bootstrapping (ICLR 2025 · similarity 0.87)
- UniBench: Visual Reasoning Requires Rethinking Vision-Language Beyond Scaling (NeurIPS 2024 · similarity 0.86)
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.
