Research map / 3D vision / 3D generation
PromptDepth: Efficient and Promptable Geometric 3D Vision Model for Embodied Intelligence
Xianyun Wang, Jiaxu Miao, Tian Xu, Siyuan Wang, Yuehao Li, Haoyang Hu, Jun Xiao, Yonghong Tian, Jun Yu
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 Gram anchoring loss to joint depth and instance training (ablation) | PromptDepth · KITTI depth, multi-task | Abs rel: 0.091 -> 0.075 | Improved |
| add Gram anchoring loss to joint training (ablation) | PromptDepth · Interactive segmentation, multi-task | mIoU: 62.4 -> 67.1 | Improved |
| add ILDS loss to joint training (ablation) | PromptDepth · KITTI depth, multi-task | Abs rel: 0.085 -> 0.075 | Improved |
| add ILDS loss to joint training (ablation) | PromptDepth · Interactive segmentation, multi-task | mIoU: 66.8 -> 67.1 | Improved |
| joint multi-task training with ILDS and Gram anchoring vs single-task depth training (ablation) | PromptDepth · KITTI depth | Abs rel: 0.081 -> 0.075 | Improved |
| joint multi-task training vs single-task segmentation training (ablation) | PromptDepth · Interactive segmentation | mIoU: 67.0 -> 67.1 | No clear effect |
| naive joint training of depth and instance tasks without Gram anchoring (ablation) | PromptDepth · KITTI depth vs single-task | delta1: 0.942 -> 0.924 | Hurt |
| train on high-fidelity post-rendered MRQ synthetic data (ablation) | PromptDepth · DAVIS-17 zero-shot tracking | J&F-Mean: 81.2 -> 83.3 | Improved |
| train on high-fidelity post-rendered MRQ synthetic data (ablation) | PromptDepth · YouTube-VOS zero-shot tracking | J&F-Mean: 75.0 -> 76.8 | Improved |
| use unified promptable model instead of SAM+DAv2 composite | embodied 3D instance-aware geometry · RTX 4090 float32 latency, mono | total latency (ms): 95.53 -> 39.09 | Improved |
| use unified promptable model instead of SAM2+VDA composite | stereo tracking · RTX 4090 float32 latency, stereo | total latency (ms): 188.46 -> 44.92 | Improved |
| train on instance-level synthetic data with unified model vs SAM | interactive segmentation by point · GraspNet | mIoU: 0.6036 -> 0.7863 | 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.
- VGGT — VGGT: Visual Geometry Grounded Transformer (CVPR 2025) · 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
- EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI (CVPR 2024 · similarity 0.91)
- BIP3D: Bridging 2D Images and 3D Perception for Embodied Intelligence (CVPR 2025 · similarity 0.88)
- Abstract 3D Perception for Spatial Intelligence in Vision-Language Models (CVPR 2026 · similarity 0.88)
- SPA: 3D Spatial-Awareness Enables Effective Embodied Representation (ICLR 2025 · similarity 0.88)
- An Embodied Generalist Agent in 3D World (ICML 2024 · similarity 0.88)
- Vision-Language Embodiment for Monocular Depth Estimation (CVPR 2025 · similarity 0.88)
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.
