Research map / 3D vision / 3D generation
Floating No More: Object-Ground Reconstruction from a Single Image
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ORG, which jointly predicts pixel height (object-ground) and perspective field (camera) maps from a single image and reprojects them into depth maps and point clouds for ground-aware object reconstruction.
Problem
Single-image 3D object reconstruction often produces objects that float or tilt when placed on a ground plane because it ignores object-ground geometry and camera parameters, which hurts 3D-aware image editing such as shadow generation and pose manipulation.
Approach
Represents the object via two dense pixel-level fields, pixel height between front/back surfaces and the ground plus a perspective field for camera parameters, and uses a perspective-guided pixel height reprojection to recover 3D point clouds without requiring known camera intrinsics or extrinsics.
Key results
- ORG achieves LSIV 1.07 and CD 0.93 for point clouds vs. best baselines (LeReS 1.11, DUSt3R 0.95) on object-centric images with off-the-shelf camera estimator.
- Joint pixel height and perspective field estimation reaches LSIV 0.81 vs. 1.25 for depth with off-the-shelf camera estimation.
- Pixel height/perspective field L1 errors of 15.4 for up-vector versus 24.5 for LeReS on the object-centric test set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ORG instead of DUSt3R with Ctrl-C | Object-centric 3D reconstruction · Object-centric test set | LSIV: 0.95 -> 0.93 | Helped |
| Use ORG instead of DUSt3R with Ctrl-C | Object-centric depth estimation · Object-centric test set | AbsRel: 17.9 -> 19.1 | Hurt |
| Use ORG instead of DUSt3R with Ctrl-C | Object-centric depth estimation · Object-centric test set | δ1: 79.5 -> 81.2 | Helped |
| Use pixel height and perspective field jointly rather than depth and off-the-shelf camera parameters (ablation) | ORG · Object-centric test set | LSIV: 1.25 -> 0.81 | Helped |
| Add perspective-field camera estimation to pixel-height reconstruction (ablation) | ORG · Object-centric test set | LSIV: 0.98 -> 0.81 | Helped |
| Use pixel height instead of depth with perspective-field camera estimation (ablation) | ORG · Object-centric test set | LSIV: 1.01 -> 0.81 | Helped |
| Use ORG rather than the baseline | Object reconstruction across viewpoint-diversity subsets · Small viewpoint diversity validation set | LSIV: 0.23 -> 0.21 | Helped |
| Use ORG rather than the baseline | Object reconstruction across viewpoint-diversity subsets · Medium viewpoint diversity validation set | LSIV: 0.37 -> 0.28 | Helped |
| Use ORG rather than the baseline | Object reconstruction across viewpoint-diversity subsets · Large viewpoint diversity validation set | LSIV: 0.72 -> 0.45 | Helped |
Surprising findings
- Pixel height representation yields better point-cloud reconstruction than depth-based learning with the same architecture and data, which the authors attribute to focusing on object-ground rather than object-camera geometry.
Limitations
- Camera-parameter-based reprojection uses a grid search over FoV and rotation, which is a heuristic step.
- Evaluation on in-the-wild data is mostly qualitative; quantitative results are on rendered Objaverse-style test data.
- Assumes a single object on a planar ground plane with the camera pointing at the object center.
Experimental setup
Datasets: Objaverse COCO (backbone pretraining)
Benchmarks: Objaverse held-out test split unseen object and human datasets
Metrics: AbsRel delta1 LSIV Chamfer distance L1 MSE
Baselines: LeReS MiDaS MegaDepth NDDepth Zero-123 DUSt3R SJC Ctrl-C
Builds on: pixel height (Zhang et al. shadow generation) perspective fields (Jin et al.) LeReS SegFormer PVTv2
Closest papers by meaning
- Orient Anything: Learning Robust Object Orientation Estimation from Rendering 3D Models (ICML 2025 · similarity 0.84)
- SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.84)
- World Tracing: Pixel-Aligned Geometry Beyond the Visible (NeurIPS 2026 · similarity 0.83)
- SeeGround: See and Ground for Zero-Shot Open-Vocabulary 3D Visual Grounding (CVPR 2025 · similarity 0.83)
- From Objects to Anywhere: A Holistic Benchmark for Multi-level Visual Grounding in 3D Scenes (NeurIPS 2025 · similarity 0.83)
- SHINOBI: Shape and Illumination using Neural Object Decomposition via BRDF Optimization In-the-wild (CVPR 2024 · similarity 0.83)
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.