Research map / Detection and segmentation / Open-set and few-shot detection
GH-ESD: Grounded Hypothesis-Driven Error Slice Discovery for Instance-Level Vision Tasks
Wei Zhang, Chaoqun Wang, Zixuan Guan, Ping Kao, Pengfei Zhao, Peng Wu, Sifeng He
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 |
|---|---|---|---|
| replace tag-based slices (HiBug2) with grounded language hypotheses and statistical verification | GH-ESD · GESD detection | Precision@10: 0.63 -> 0.73 | Improved |
| use same Gemini backend for HiBug2 as GH-ESD (ablation) | HiBug2 · GESD detection | Precision@10: 0.63 -> 0.63 | No clear effect |
| add data-driven hypotheses to knowledge-driven ones (ablation) | GH-ESD · GESD detection | Precision@10: 0.62 -> 0.73 | Improved |
| use instance-level grounding instead of image-level aggregation (ablation) | GH-ESD · faces partially hidden by objects | Precision@10: 0.60 -> 0.80 | Improved |
| use trend-based verification instead of error-rate thresholding (ablation) | GH-ESD · GESD | F1: average 3.7% F1 gain | Improved |
| use Spearman correlation test instead of max-slope trend (ablation) | GH-ESD · GESD | F1: 0.77 -> 0.78 | No clear effect |
| use yes/no logits instead of direct numeric score (ablation) | GH-ESD · GESD subset | mAP: 0.23 -> 0.47 | Improved |
| use yes/no logits instead of yes/no/not-sure categorical (ablation) | GH-ESD · GESD subset | mAP: 0.18 -> 0.47 | Improved |
| use Qwen2.5-VL-7B instead of SigLIP for matching (ablation) | GH-ESD · NICO++ (95%) | Precision@10: 0.51 -> 0.69 | Improved |
| use GPT-4o instead of Qwen2.5-VL-7B for matching (ablation) | GH-ESD · NICO++ (95%) | Precision@10: 0.69 -> 0.71 | No clear effect |
| swap hypothesis LLM from Gemini-2.5-Pro to GPT-4o or Claude-3.5-Sonnet (ablation) | GH-ESD · NICO++ (95%) | Precision@10: 0.69 vs 0.68 vs 0.67 | No clear effect |
| slice-aware GroupDRO repair using GH-ESD slice | Faster R-CNN · COCO bicycle | mAP-bicycle: 28.58 -> 33.59 | Improved |
Closest papers by meaning
- HiBug2: Efficient and Interpretable Error Slice Discovery for Comprehensive Model Debugging (ICLR 2025 · similarity 0.87)
- Reflexive Guidance: Improving OoDD in Vision-Language Models via Self-Guided Image-Adaptive Concept Generation (ICLR 2025 · similarity 0.86)
- $\Delta \mathrm{Energy}$: Optimizing Energy Change During Vision-Language Alignment Improves both OOD Detection and OOD Generalization (NeurIPS 2025 · similarity 0.85)
- COUNTS: Benchmarking Object Detectors and Multimodal Large Language Models under Distribution Shifts (CVPR 2025 · similarity 0.85)
- Generalize or Detect? Towards Robust Semantic Segmentation Under Multiple Distribution Shifts (NeurIPS 2024 · similarity 0.85)
- GCD: Advancing Vision-Language Models for Incremental Object Detection via Global Alignment and Correspondence Distillation (AAAI 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.
