SparseCraft: Few-Shot Neural Reconstruction through Stereopsis Guided Geometric Linearization (2407.14257v1)

Abstract: We present a novel approach for recovering 3D shape and view dependent appearance from a few colored images, enabling efficient 3D reconstruction and novel view synthesis. Our method learns an implicit neural representation in the form of a Signed Distance Function (SDF) and a radiance field. The model is trained progressively through ray marching enabled volumetric rendering, and regularized with learning-free multi-view stereo (MVS) cues. Key to our contribution is a novel implicit neural shape function learning strategy that encourages our SDF field to be as linear as possible near the level-set, hence robustifying the training against noise emanating from the supervision and regularization signals. Without using any pretrained priors, our method, called SparseCraft, achieves state-of-the-art performances both in novel-view synthesis and reconstruction from sparse views in standard benchmarks, while requiring less than 10 minutes for training.

• The paper introduces a novel method that leverages implicit neural representations regularized by MVS cues to eliminate the need for pre-trained data.
• It integrates a progressive multi-resolution hash encoding strategy with Taylor expansion losses to boost efficiency and robustness.
• SparseCraft achieves state-of-the-art results on benchmarks by delivering accurate geometric reconstructions and exceptional photometric fidelity.

SparseCraft: Few-Shot Neural Reconstruction through Stereopsis Guided Geometric Linearization

Mae Younes, Amine Ouasfi, and Adnane Boukhayma present a method in their paper titled "SparseCraft: Few-Shot Neural Reconstruction through Stereopsis Guided Geometric Linearization." This research proposes a novel technique for 3D shape reconstruction and novel view synthesis from a minimal number of color images, emphasizing efficiency and the elimination of the need for pre-trained data priors. SparseCraft fundamentally addresses a significant challenge in the field by combining multi-view stereo (MVS) cues with an implicit neural representation approach.

At its core, SparseCraft operates by training an implicit neural model to capture both the shape and radiance of objects given sparse view inputs. The primary components of this model include a Signed Distance Function (SDF) and a radiance field. This model is incrementally trained using ray marching-enabled volumetric rendering and is regularized by stereo cues derived from MVS. Noteworthy is the absence of any reliance on pre-trained data priors, which distinguishes this approach from several contemporary methods that use voluminous and costly pre-training databases.

Key contributions of this work include:

1. Few-shot 3D Reconstructions and Novel View Synthesis: SparseCraft achieves state-of-the-art performance in both 3D reconstruction and novel view synthesis from a sparse set of images without leveraging pre-learned data priors.
2. Progressive Multi-Resolution Hash Learning Strategy: The method utilizes a progressive learning strategy with multi-resolution hash encoding, enhancing the efficiency and stability of training.
3. Integration of MVS Data: The framework uniquely integrates all geometric and radiometric information from MVS, including points, normals, and color, to regularize the SDF and diffuse components.
4. Taylor Expansion Inspired Losses: To mitigate the influence of noise inherently present in sparse data and MVS, the paper introduces novel Taylor expansion based loss functions for better accuracy and robustness.

In experimental evaluations, SparseCraft demonstrates superior performance across multiple benchmark datasets, including DTU, BlendedMVS, and Tanks and Temples. The method not only produces accurate metric 3D reconstructions but also generates photo-realistic novel views. For instance, SparseCraft achieves an average Chamfer distance of 0.01, PSNR of 20.55, and LPIPS of 0.116 when evaluated on the DTU dataset.

The robust performance of SparseCraft is attributed to several key design choices. First, regularizing the SDF near the surface improves generalization and limits overfitting to noise. Second, incorporating MVS-derived normals and color information enriches the training signal, enhancing both geometric and photometric fidelity. Third, the progressive hash encoding strategy prevents premature overfitting to fine details, a common issue in sparse settings.

Practically, SparseCraft's significant reduction in required input data widens the applicability of 3D reconstruction to scenarios with stringent data constraints, such as low-budget or out-of-studio settings. Theoretically, this approach advances the understanding of integrating photogrammetry with neural implicit representations, paving the way for future work to explore other forms of sparse supervision or hybridized learning strategies in 3D vision tasks.

In conclusion, SparseCraft presents a compelling advancement in few-shot neural reconstruction, addressing both practical constraints and theoretical challenges with innovative strategies. Future work may explore extending SparseCraft's framework to other forms of sparsity or investigating its application to dynamic scenes, augmenting possibilities in neural scene representation and rendering.