Camera3DMM: Leveraging Perspective Camera for Estimating Parametric 3D Head Models

Vishwesh Vhavle1,2, Hiteshi Jain1, Avinash Sharma2
1Mercedes-Benz R&D India, 2IIT Jodhpur
SIGGRAPH Asia 2025 Technical Communications
Paper Supplementary Material
Camera3DMM Teaser

Camera3DMM enables 3D reconstruction of human heads under perspective distortions. We visualize selfie image samples from the NoW dataset on top, and our FLAME fitting results on bottom.

Abstract

3D human head modeling is often formulated under scaled-orthographic assumptions, which fail in close-range scenarios such as handheld mobile device captures, where perspective distortion dominates and leads to unstable and inconsistent reconstructions.

We propose Camera3DMM, a novel perspective-aware 3D human head reconstruction framework that jointly estimates 3D facial geometry and camera parameters from a single image. To address the lack of perspective-rich training data, we leverage high-quality 3D RGB scans to render images with pseudo ground truth labels across diverse focal lengths and perspective distortions, thereby enabling explicit modeling of perspective variability.

Trained on this data, Camera3DMM achieves stable and consistent reconstructions under varying intrinsics and demonstrates a 22% improvement in mesh quality over the best-performing baseline. These results establish Camera3DMM as a strong baseline for perspective-aware 3D face reconstruction, particularly in challenging close-range scenarios.

Method Overview

Method Overview

Overview of the proposed method for 3D head reconstruction: (1) Data Generation, (2) Data Transform, (3) Network Training, and (4) Inference with Optimization.

Results

Qualitative Results

FLAME meshes predicted by DECA, EMOCAv2, SMIRK, and Camera3DMM (Ours) against ground truth. Our method exhibits more accurate facial alignment across different focal lengths.

Qualitative Comparisons along with corresponding Vertex error of FLAME meshes predicted by DECA, EMOCAv2, SMIRK, and Camera3DMM (Ours) against ground truth.

BibTeX

@article{vhavle2025camera3dmm,
  title={Camera3DMM: Leveraging Perspective Camera for Estimating Parametric 3D Head Models},
  author={Vhavle, Vishwesh and Jain, Hiteshi and Sharma, Avinash},
  journal={SIGGRAPH Asia 2025 Technical Communications},
  year={2025}
}