Single-view Textured Human Reconstruction with Image-Conditioned Diffusion

Concepts de base

SiTH is a novel pipeline that integrates an image-conditioned diffusion model to reconstruct high-quality and fully textured 3D human meshes from single images.

Résumé

The paper proposes SiTH, a two-stage pipeline for single-view textured human reconstruction.

In the first stage, SiTH employs an image-conditioned diffusion model to hallucinate perceptually consistent back-view appearances based on the input front-view image. This is achieved by adapting the network architecture of a pretrained diffusion model to enable conditioning on the front-view image, UV map, and silhouette mask.

In the second stage, SiTH leverages the input front-view image and the generated back-view image to reconstruct a full-body textured mesh. It utilizes a skinned body prior to handle the 3D ambiguity and employs pixel-aligned feature querying to learn a mapping from the images to the 3D geometry and texture.

The authors demonstrate that SiTH can efficiently produce high-quality and diverse 3D textured humans from single images, including unseen photos and AI-generated images. Extensive evaluations on two 3D human benchmarks show that SiTH outperforms state-of-the-art methods in terms of accuracy and perceptual quality.

Stats

"A long-standing goal of 3D human reconstruction is to create lifelike and fully detailed 3D humans from single-view images."
"SiTH requires as few as 500 3D human scans for training while maintaining its generality and robustness to diverse images."
"Compared to existing end-to-end methods, our two-stage pipeline can recover full-body textured meshes, including back-view details, and demonstrates robustness to unseen images."
"In contrast to time-intensive diffusion-based optimization methods, our pipeline efficiently produces high-quality textured meshes in under two minutes."

Citations

"To address the above challenges, we propose SiTH, a novel pipeline that integrates an image-conditioned diffusion model to reconstruct lifelike 3D textured humans from monocular images."
"Notably, both modules in the pipeline can be efficiently trained with 500 textured human scans in THuman2.0 [75] and generalize unseen images."

Idées clés tirées de

SiTH

by Hsuan-I Ho,J... à arxiv.org 04-02-2024

https://arxiv.org/pdf/2311.15855.pdf

Questions plus approfondies

How can the proposed pipeline be extended to handle dynamic human motions and animations

To extend the proposed pipeline to handle dynamic human motions and animations, we can incorporate techniques from the field of computer animation. One approach could be to utilize motion capture data to drive the animations of the reconstructed 3D human models. By capturing the movements of real humans and mapping them onto the reconstructed models, we can achieve realistic and dynamic animations. Additionally, techniques such as skeletal animation can be employed to animate the 3D models based on the underlying skeleton structure. This would involve defining key poses and transitions between them to create fluid and natural movements in the animations. Furthermore, incorporating physics-based simulations for cloth and hair dynamics can add another layer of realism to the animations, making them more lifelike and engaging.

What are the potential limitations of the image-conditioned diffusion model in handling extreme poses or occlusions

While the image-conditioned diffusion model is powerful in generating realistic back-view images for 3D reconstruction, it may face limitations when handling extreme poses or occlusions. Extreme poses can lead to distortions in the generated images, affecting the accuracy of the 3D reconstruction. Similarly, occlusions in the input images can result in missing or incorrect information in the generated back-view images, impacting the overall quality of the reconstructed 3D models. To address these limitations, additional training data with a diverse range of poses and occlusions can be used to improve the model's robustness. Augmenting the training data with synthetic examples of extreme poses and occlusions can help the model learn to handle such scenarios better. Moreover, incorporating advanced image processing techniques, such as inpainting and completion algorithms, can help fill in missing information in occluded regions, enhancing the model's performance in challenging situations.

How can the 3D reconstruction quality be further improved by incorporating additional sensor data, such as depth or multi-view images

To further improve the quality of 3D reconstruction, incorporating additional sensor data such as depth or multi-view images can provide valuable information for enhancing the accuracy and detail of the reconstructed models. Depth data can help in capturing the spatial information of the scene, enabling more precise reconstruction of the 3D geometry. By integrating depth information into the pipeline, we can improve the depth estimation and surface reconstruction of the 3D models. Multi-view images, obtained from different viewpoints, can offer more comprehensive coverage of the subject, reducing ambiguities in the reconstruction process. Utilizing multi-view geometry reconstruction techniques, such as structure-from-motion or multi-view stereo, can help in generating more detailed and accurate 3D models by leveraging information from multiple perspectives. Additionally, sensor fusion techniques can be employed to combine data from different sensors, such as RGB cameras and depth sensors, to enhance the overall quality and fidelity of the reconstructed 3D models.