Metric Depth Estimation on DepthVLM-Bench 1.0 (test)
81Delta1 Accuracy (Argoverse2)Ours-4B
Evaluation results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Ours-4BCategory=VLMs Trained on Metric Depth Estimation, Backbone=Qwen3-VL-4B2026.05 | 81 | 87.9 | 81.8 | 82.1 | 92.4 | 86.1 | 88.2 | 91.2 | 90.8 | 86.8 | |
| Ours-8BCategory=VLMs Trained on Metric Depth Estimation, Backbone=Qwen3-VL-8B2026.05 | 79.8 | 86.5 | 81.3 | 83.1 | 92.8 | 90.1 | 88.9 | 93.6 | 92 | 87.6 | |
| DepthLM-12BCategory=VLMs Trained on Metric Depth Estimation2026.05 | 76.1 | 58.8 | 65.4 | 73.6 | 66.6 | 75.6 | 78.5 | 75.4 | 86.6 | 73 | |
| Youtu-VL-4BCategory=VLMs Trained on Metric Depth Estimation2026.05 | 66.3 | 47.3 | 34.2 | 69.8 | 28.6 | 52.2 | 73.4 | 85.6 | 84.9 | 60.3 | |
| GPT-5.5Category=General-Purpose VLMs, Prompting Strategy=DepthLM [9]2026.05 | 37.8 | 36.8 | 30.4 | 27.6 | 36.9 | 43.2 | 52.7 | 48.3 | 52.5 | 40.7 | |
| GPT-4oCategory=General-Purpose VLMs, Prompting Strategy=DepthLM [9]2026.05 | 14.1 | 13.9 | 19.3 | 17.4 | 29.6 | 36.9 | 35.8 | 39.3 | 39.4 | 27.3 | |
| InternVL3.5-8BCategory=General-Purpose VLMs, Prompting Strategy=DepthLM [9]2026.05 | 13.9 | 11 | 17.7 | 11.1 | 18.5 | 39.5 | 45.9 | 21.4 | 43.1 | 24.7 | |
| InternVL3.5-38BCategory=General-Purpose VLMs, Prompting Strategy=DepthLM [9]2026.05 | 13.1 | 12.9 | 21.9 | 11 | 18.1 | 40 | 43.1 | 15.5 | 42.3 | 24.2 | |
| Qwen3-VL-8BCategory=General-Purpose VLMs, Prompting Strategy=DepthLM [9]2026.05 | 11.9 | 6.1 | 24.3 | 11.8 | 17.5 | 27.4 | 28.6 | 8.1 | 19.4 | 17.2 | |
| InternVL3.5-14BCategory=General-Purpose VLMs, Prompting Strategy=DepthLM [9]2026.05 | 8.6 | 4.5 | 12.8 | 8.4 | 18.3 | 39 | 44 | 23.5 | 44.5 | 22.6 | |
| SpatialRGPT-8BCategory=Spatial-Enhanced VLMs, Prompting Strategy=DepthLM [9]2026.05 | 4.5 | 6.4 | 9.6 | 11.6 | 13.3 | 12.4 | 8.4 | 4.4 | 7 | 8.6 | |
| Qwen3-VL-4BCategory=General-Purpose VLMs, Prompting Strategy=DepthLM [9]2026.05 | 3.3 | 0.6 | 8.6 | 4 | 8.5 | 20.8 | 22 | 3.7 | 15.5 | 9.7 | |
| Qwen3-VL-32BCategory=General-Purpose VLMs, Prompting Strategy=DepthLM [9]2026.05 | 2.9 | 1.7 | 13.8 | 4.9 | 15 | 43.4 | 53.6 | 10.5 | 43.5 | 21 | |
| SpaceLLaVA-13BCategory=Spatial-Enhanced VLMs, Prompting Strategy=DepthLM [9]2026.05 | 0.6 | 0.2 | 0.1 | 0.6 | 10.7 | 5 | 4.4 | 17.2 | 8.7 | 5.3 | |
| Cambrian-S-7BCategory=Spatial-Enhanced VLMs, Prompting Strategy=DepthLM [9]2026.05 | 0.6 | 1.9 | 3.8 | 3.3 | 6.9 | 14.5 | 7.3 | 5.7 | 6.3 | 5.6 | |
| Always Output 2.0mCategory=Naive Prediction with Constant Answers2026.05 | 0.2 | 0.4 | 0.2 | 1 | 11.2 | 26.1 | 38 | 26.9 | 37.3 | 15.7 |