Research map / 3D vision / Human pose estimation

ODHSR: Online Dense 3D Reconstruction of Humans and Scenes from Monocular Videos

CVPR 2025 · 3D vision › Human pose estimation · approach: Transformer-based multi-view 3D pose · 4 citations (OpenAlex)

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.

Presents ODHSR, an online Gaussian-splatting SLAM framework that jointly tracks camera and human poses and reconstructs photorealistic human avatars and scenes from a monocular RGB video.

Problem

Reconstructing photorealistic human and scene jointly from monocular in-the-wild video typically needs pre-calibrated cameras, known human poses, and days of offline training; online joint camera tracking, human pose estimation and dense human-scene reconstruction is lacking.

Approach

Extends RGB-only Gaussian Splatting SLAM (GS-ICP/Gaussian SLAM style, based on [41]) with a holistic human-scene 3D Gaussian representation: SMPL-guided canonical human Gaussians with hash-encoded time-pose non-rigid deformation, occlusion-aware human silhouette rendering, and monocular depth priors, jointly optimized for camera and human pose during tracking and mapping.

Key results

• On EMDB, full ODHSR reaches 23.790 PSNR whole-image and 28.955 human-only PSNR, outperforming HUGS (21.605 / 26.165) and HSR (18.675 / 25.127).
• Camera tracking ATE of 8.4 cm on EMDB, comparable to DROID-SLAM, and WA-MPJPE of 175.215 mm versus 636.001 mm for WHAM.
• Reported as about 75x faster than HSR, with 85 FPS rendering-oriented training throughput per image on a single RTX 4090 (training FPS as listed in Table 1).

Measured effects reported in the paper

Surprising findings

• Ablation shows removing the optical flow loss raises ATE error from 0.084 to 0.214 m, indicating tracking is highly sensitive to flow supervision.
• Human-only GauHuman and 3DGS-Avatar outperform ODHSR in human-only PSNR on NeuMan when given ground-truth camera poses, with ODHSR's human-only quality second only to 3DGS-Avatar.

Limitations

• Relies on pre-estimated human segmentation, optical flow and monocular depth priors, and an off-the-shelf pose initializer (WHAM).
• Human-only quality drops on NeuMan when switching from ground-truth to estimated camera poses.

Experimental setup

Datasets: EMDB NeuMan

Benchmarks: EMDB NeuMan

Metrics: PSNR SSIM LPIPS ATE RMSE MPJPE PA-MPJPE MVE W-MPJPE WA-MPJPE jitter

Baselines: HSR Vid2Avatar HUGS 3DGS-Avatar GauHuman DROID-SLAM WHAM

Builds on: Gaussian Splatting SLAM 3DGS SMPL Splat-SLAM WHAM

Closest papers by meaning

• Guess The Unseen: Dynamic 3D Scene Reconstruction from Partial 2D Glimpses (CVPR 2024 · similarity 0.91)
• Reconstructing People, Places, and Cameras (CVPR 2025 · similarity 0.91)
• UniCon3R: Unified Contact-aware 4D Human-Scene Reconstruction from Monocular Video (NeurIPS 2026 · similarity 0.89)
• Link to the Past: Temporal Propagation for Fast 3D Human Reconstruction from Monocular Video (CVPR 2025 · similarity 0.89)
• Synergistic Global-space Camera and Human Reconstruction from Videos (CVPR 2024 · similarity 0.89)
• No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images (ICLR 2025 · similarity 0.89)