Generative approaches like Point-E and Shap-E generate sharper surfaces but often produce details that are hallucinated and do not accurately follow the input images.
Previous regression-based methods, such as MCC, adhere to input cues effectively, but their depiction of occluded surfaces may still be inaccurate.
ZeroShape reconstruction not only captures the global shape structure but also faithfully adheres to local geometry cues from the input image, surpassing prior methods.
Our ablation study shows that incorporating explicit geometric reasoning significantly enhances model performance, indicating its importance in achieving accurate results.
Collection
[
|
...
]