We now present our architecture for shape reconstruction. Our architecture is based on two established practices from prior works in this field

This article presents a new way to predict the 3D shape of an object from a single image.

We show that ZeroShape not only achieves superior performance over state-of-the-art methods.

We use three different real-world dataset evaluation: OmniObject3D, Ocrtoc3D and Pix3D.

We use all 55 categories of ShapeNetCore.v2 and over 1000 categories from the Objaverse-LVIS subset.

We compare our approach to other state-of-the-art methods on the benchmark we curated. We now present and analyze the quantitative results for each dataset.

We consider five state-of-the-art baselines for shape reconstruction, SS3D, MCC, Point-E, Shap-E and OpenLRM.

We present a strong regression-based model for zero-shot shape reconstruction. The core of our model is an intermediate representation of the 3D visible surface

Due to computational resource limitations, we are not able to process and train our model on the full Objaverse dataset.

We show qualitative results of different methods in Fig. 5. Generative approaches such as Point-E and Shap-E tend to have sharper surfaces

In this section, we describe our data generation procedure for training and rendering the object scans from OmniObject3D

To test the out-of-domain generalization ability, we generate images of imaginary objects as the input to our model (see Fig. 10).

We show additional qualitative results on OmniObject3D, Ocrtoc3D, and Pix3D in Fig. 7, Fig. 8 and Fig. 9, respectively.