Invisible Stitch: Generating Smooth 3D Scenes with Depth Inpainting

TL;DR

A depth completion model is introduced to directly learn the 3D fusion process, resulting in improved geometric coherence of generated scenes and a new benchmark to evaluate the geometric accuracy of scene generation methods is introduced.

Abstract

3D scene generation has quickly become a challenging new research direction, fueled by consistent improvements of 2D generative diffusion models. Most prior work in this area generates scenes by iteratively stitching newly generated frames with existing geometry. These works often depend on pre-trained monocular depth estimators to lift the generated images into 3D, fusing them with the existing scene representation. These approaches are then often evaluated via a text metric, measuring the similarity between the generated images and a given text prompt. In this work, we make two fundamental contributions to the field of 3D scene generation. First, we note that lifting images to 3D with a monocular depth estimation model is suboptimal as it ignores the geometry of the existing scene. We thus introduce a novel depth completion model, trained via teacher distillation and self-training to learn the 3D fusion process, resulting in improved geometric coherence of the scene. Second, we introduce a new benchmarking scheme for scene generation methods that is based on ground truth geometry, and thus measures the quality of the structure of the scene.

Figures (4)

• Figure 1: Overview of our 3D scene generation method. Starting from an input image $I_0$, we project it to a point cloud based on a depth map predicted by a depth estimation network $g$. To extend the scene, we render it from a new view point and query a generative model $f$ to hallucinate beyond the scene's boundary. Now, we condition $g$ on the depth of the existing scene and the image of the scene extended by $f$ to produce a geometrically consistent depth map to project the hallucinated points. This process may be repeated until a 360-degree scene has been generated.
• Figure 2: Overview of our training procedure. In this compact training scheme a depth completion network $g$ is learned by jointly training depth inpainting as well as depth prediction without a sparse depth input (the ratio being determined by the task probability $p$). A teacher network $g_T$ is utilized to generate a pseudo ground-truth depth map $D$ for a given image $I$. This depth map is then masked with a random mask $M$, to obtain a sparse depth input $\tilde{D}$.
• Figure 3: Overview of our scene consistency evaluation approach. Assume a scene is described by a set of views $\{v_1, v_2, \dots\}$ with associated images, depth maps, and camera poses, where the overlap of two views is described by a function $\phi(v_i, v_j)$. For a given view pair $(v_i, v_j)$ with $\phi(v_i, v_j) \geq \tau$, we generate a representation, e.g., a point cloud, from the ground-truth (GT) data for $v_i$. Then, we render the representation from the view point of $v_j$. We feed the corresponding ground-truth image and the representation's depth into the model under consideration to extrapolate the missing depth. Finally, we calculate the mean absolute error between the result and the ground-truth depth for $v_j$, only considering those regions that were extrapolated.
• Figure 4: Qualitative results of our method on real-world images. We show hallucinated views and the corresponding depth maps of 360-degree scenes. We also provide a full view of the generated 360-degree scene as well as a more detailed cut-away view.