Research map / Detection and segmentation / Open-set and few-shot detection
Conversational Image Segmentation: Grounding Abstract Concepts with Scalable Supervision
Aadarsh Sahoo, Georgia Gkioxari
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 conversational post-training (Phase 2) on data-engine data (ablation) | ConverSeg-Net 3B · ConverSeg SAM-seeded (All) | gIoU: 58.8 -> 70.5 | Improved |
| add conversational post-training (Phase 2) (ablation) | ConverSeg-Net 3B · ConverSeg SAM-seeded (Physics & Safety) | gIoU: 43.8 -> 60.7 | Improved |
| scale prompt encoder from Qwen2.5-VL-3B to 7B (ablation) | ConverSeg-Net · ConverSeg SAM-seeded (All) | gIoU: 70.5 -> 73.3 | Improved |
| add LoRA finetuning of the Qwen prompt encoder (vs frozen) (ablation) | ConverSeg-Net · ConverSeg human-annotated (All) | gIoU: 49.4 -> 64.4 | Improved |
| give the prompt encoder image plus text instead of text only (ablation) | ConverSeg-Net · ConverSeg human-annotated (All) | gIoU: 47.4 -> 64.4 | Improved |
| add dense EOS prompt embedding to sparse embeddings (ablation) | ConverSeg-Net · ConverSeg human-annotated (All) | gIoU: 63.9 -> 64.4 | Improved |
| swap Qwen2.5-VL-3B for Perception-LM-3B backbone (ablation) | ConverSeg-Net · ConverSeg human-annotated (All) | gIoU: 64.4 -> 65.2 | No clear effect |
| train only on conversational data (no curriculum) (ablation) | ConverSeg-Net · RefCOCO/+/g (avg over 9 splits) | gIoU: 76.2 -> 68.0 | Hurt |
| joint training on all data without curriculum instead of full curriculum (ablation) | ConverSeg-Net · ConverSeg human-annotated (All) | gIoU: 64.4 -> 61.9 | Hurt |
| add 50-50 data mix in Phase 2 (vs conversational-only Phase 2) (ablation) | ConverSeg-Net · RefCOCO/+/g (avg over 9 splits) | gIoU: 74.1 -> 76.2 | Improved |
| add conversational Phase 2 on top of Phase 1 (ablation) | ConverSeg-Net · ConverSeg human-annotated (All) | gIoU: 56.4 -> 64.4 | Improved |
| add conversational post-training (zero-shot transfer to ReasonSeg) (ablation) | ConverSeg-Net 3B · ReasonSeg test (zero-shot) | gIoU: 49.9 -> 55.1 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- X-SAM — X-SAM: From Segment Anything to Any Segmentation (AAAI 2026) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PhraseCut: Language-Based Image Segmentation in the Wild (CVPR 2020 · similarity 0.89)
- Mask Grounding for Referring Image Segmentation (CVPR 2024 · similarity 0.89)
- SetCon: Towards Open-Ended Referring Segmentation via Set-Level Concept Prediction (NeurIPS 2026 · similarity 0.87)
- Emergent Open-Vocabulary Semantic Segmentation from Off-the-shelf Vision-Language Models (CVPR 2024 · similarity 0.87)
- Early Semantic Grounding in Image Editing Models for Zero-Shot Referring Image Segmentation (NeurIPS 2026 · similarity 0.87)
- Seg-ReSearch: Segmentation with Interleaved Reasoning and External Search (ICML 2026 · similarity 0.87)
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.
