OAHuman: Occlusion-Aware 3D Human Reconstruction from Monocular Images
Authors: Yuanwang Yang, Hongliang Liu, Muxin Zhang, Nan Ma, Jingyu Yang, Yu-Kun Lai, Kun Li
Organizations: College of Intelligence and Computing, Tianjin University, Tianjin 300350, China · College of Artificial Intelligence, Beijing University of Technology, Beijing 100124, China · School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China · School of Computer Science and Informatics, Cardiff University, Cardiff CF24 4AG, United Kingdom
Monocular 3D human reconstruction in real-world scenarios remains highly challenging due to frequent occlusions from surrounding objects, people, or image truncation. Such occlusions lead to missing geometry and unreliable appearance cues, severely degrading the completeness and realism of reconstructed human models. Although recent neural implicit methods achieve impressive results on clean inputs, they struggle under occlusion due to entangled modeling of shape and texture. In this paper, we propose OAHuman, an occlusion-aware framework that explicitly decouples geometry reconstruction and texture synthesis for robust 3D human modeling from a single RGB image. The core innovation lies in the decoupling-perception paradigm, which addresses the fundamental issue of geometry-texture cross-contamination in occluded regions. Our framework ensures that geometry reconstruction is perceptually reinforced even in occluded areas, isolating it from texture interference. In parallel, texture synthesis is learned exclusively from visible regions, preventing texture errors from being transferred to the occluded areas. This decoupling approach enables OAHuman to achieve robust and high-fidelity reconstruction under occlusion, which has been a long-standing challenge in the field. Extensive experiments on occlusion-rich benchmarks demonstrate that OAHuman achieves superior performance in terms of structural completeness, surface detail, and texture realism, significantly improving monocular 3D human reconstruction under occlusion conditions.
Figures & tables
Fig. 1: OAHuman enables robust 3D human reconstruction under diverse occlusions. Our method successfully reconstructs complete geometry and realistic textures from monocular images across challenging occlusion types, including human-induced (left), object-induced (middle), and in-the-wild scenarios (right). Input highlights are added for visualization to indicate the target subject.
Fig. 2: Overview of OAHuman. Our method consists of a two-stage pipeline for occlusion-aware 3D human reconstruction. The first stage performs geometry restoration using a visibility-guided coarse completion (VGCC) module with feature-level supervisor regularization (FLSR) and dual-view normal refinement (DVNR). The second stage synthesizes semantically consistent textures via a geometry-faithful diffusion renderer, enabling high-fidelity reconstruction under occlusion.
Fig. 3: Feature-level supervisor regularization. We use a supervisor network trained on unoccluded data to provide hierarchical feature guidance for geometry completion. We achieve emphasis on global structure and suppression of low-level noise through depth-aware feature weighting, and enhance supervision on occluded body regions through a visibility-aware masking strategy.
Fig. 4: Qualitative comparison of geometry reconstruction results.
Fig. 5: Qualitative comparison of novel view synthesis results.
Method
CustomHumans [ 33 ]
2k2k [ 34 ]
Geometry
Novel View Synthesis
Geometry
Novel View Synthesis
CD ↓
P2S ↓
Normal ↓
SSIM ↑
PSNR ↑
LPIPS ↓
CD ↓
P2S ↓
Normal ↓
SSIM ↑
PSNR ↑
LPIPS ↓
ECON
1.9223
1.2334
1.5150
-
-
-
2.8937
2.6024
3.1966
-
-
-
SiTH
4.2868
1.8328
2.0597
0.8698
23.08
0.1261
4.1096
3.2313
3.0603
0.8706
15.07
0.1448
R 2 Human
1.6711
1.5694
1.6085
0.8331
21.84
0.1981
2.9063
2.5428
2.4865
0.8325
15.53
0.1168
PSHuman
3.0263
2.5027
2.7213
0.8257
18.85
0.4894
3.2485
2.6949
2.7534
0.8838
15.99
0.1142
TABLE I: Quantitative comparison of geometry reconstruction and novel view synthesis.
Method
CD ↓
P2S ↓
Normal ↓
VGCC w/o FLSR
1.4283
1.4281
1.4551
VGCC
1.2547
1.2275
1.3717
2D Normal Completion
0.9835
0.9408
1.2124
Ours(full)
0.9440
0.8831
1.1082
TABLE II: Quantitative results of ablation study.
Fig. 6: Qualitative results of ablation study.
Method
SSIM ↑
PSNR ↑
LPIPS ↓
w/o Mesh normal
0.8672
22.45
0.0891
w/ Mesh normal
0.8759
23.51
0.0833
TABLE III: Ablation on the renderer with different normal sources.
Fig. 7: Qualitative comparison of renderer ablations. Using refined mesh normals leads to more faithful texture generation.
Fig. 8: Qualitative reconstruction results under accumulative occlusion.
Dept. of Electrical and Computer Engineering, Seoul National University, Republic of Korea · IPAI, Seoul National University, Republic of Korea · INMC, Seoul National University, Republic of Korea