3D Visual Grounding on ScanRefer Multiple
52.5Accuracy @ IoU=0.50TAB
Evaluation results
| Method | Links | ||
|---|---|---|---|
| TABLLM/VLM=Qwen3-VL-32B, w/o PC=✗2026.04 | 52.5 | 60.8 | |
| GPT4SceneLLM/VLM=Qwen2-VL-7B, w/o PC=✗2026.04 | 50.9 | 56.4 | |
| 3D-RSLLM/VLM=LLaVA-Next-Video, w/o PC=✗2026.04 | 50.8 | 57 | |
| CAPrunerBackbone=Llama-3-8B2026.04 | 49.9 | 55.3 | |
| 3DGraphLLMSize=8B2026.04 | 49.4 | 54.7 | |
| QuatRoPE2026.04 | 49.2 | 54.3 | |
| PQ3D2026.04 | 46.2 | — | |
| Video-3D-LLMLLM/VLM=LLaVA-Video-7B, w/o PC=✗2026.04 | 45.3 | 50.9 | |
| SPAZERLLM/VLM=GPT-4o, w/o PC=✗2026.04 | 43.4 | 51.7 | |
| Chat-SceneLLM/VLM=Vicuna-7B, w/o PC=✗2026.04 | 42.9 | 47.7 | |
| CAPrunerBackbone=Llama-3.2-1B2026.04 | 42.8 | 48 | |
| MA2TransVG2026.04 | 41.4 | 53.8 | |
| SeqVLMLLM/VLM=Doubao-1.5-pro, w/o PC=✗2026.04 | 41.3 | 47.8 | |
| 3DGraphLLMSize=1B2026.04 | 40.5 | 45 | |
| TABLLM/VLM=Qwen3-VL-32B, w/o PC=✓2026.04 | 39.9 | 60.1 | |
| G3-LQLLM/VLM=-, w/o PC=✗2026.04 | 39.7 | 50.2 | |
| AugRefer2026.04 | 39.1 | 50 | |
| 3D-VisTA2026.04 | 39.1 | 43.7 | |
| EDALLM/VLM=-, w/o PC=✗2026.04 | 37.6 | 49.1 | |
| BUTD-DETRLLM/VLM=-, w/o PC=✗2026.04 | 35.1 | 46.6 | |
| VLM-GrounderLLM/VLM=GPT-4o, w/o PC=✓2026.04 | 33.5 | 48.3 | |
| SeeGroundLLM/VLM=Qwen2-VL-72B, w/o PC=✗2026.04 | 30 | 34 | |
| SeeGround2026.04 | 30 | 34 | |
| 3DVG-TLLM/VLM=-, w/o PC=✗2026.04 | 28.7 | 38.4 | |
| CSVGLLM/VLM=Mistral-Large-2407, w/o PC=✗2026.04 | 27.3 | 38.4 | |
| CSVG2026.04 | 27.3 | 38.4 | |
| FFL-3DOG2026.04 | 25.7 | 35.2 | |
| ZSVG3DLLM/VLM=GPT-4 turbo, w/o PC=✗2026.04 | 24.6 | 27.7 | |
| ScanReferLLM/VLM=-, w/o PC=✗2026.04 | 21.3 | 32.1 |