Lightweight Neural Framework for Robust 3D Volume and Surface Estimation from Multi-View Images

Estimating the volume and surface area of 3D objects from 2D images is essential for fields like marine ecology, medical diagnostics, and dietary analysis. Traditional methods often rely on dense multi-view stereo or explicit 3D mesh generation, which are computationally intensive, require high-quality data, and involve complex post-processing. This paper asks whether a lightweight, end-to-end neural framework can accurately predict these geometric metrics directly from sparse, noisy multi-view images without requiring iterative optimization or watertight mesh reconstruction.

The proposed framework utilizes a multi-modal architecture that integrates 2D and 3D information. First, it uses promptable segmentation tools to isolate the object of interest. It then processes the input images through two parallel branches: a 3D branch that generates a fused point cloud and a 2D branch that extracts visual features using a pre-trained DINOv3 encoder. These representations are fused and fed into a graph-based decoder that regresses the volume and surface area. Crucially, the model employs a Deep Evidential Regression objective, which allows it to output not just point estimates, but also calibrated measures of aleatoric and epistemic uncertainty.

The framework demonstrates superior accuracy and robustness compared to state-of-the-art methods, particularly in scenarios with sparse input views. By avoiding explicit mesh generation, the system achieves rapid inference, making it suitable for large-scale surveys and resource-constrained environments. The inclusion of the 2D feature branch significantly improves estimation accuracy, while the evidential uncertainty estimation provides a reliable metric for the model's confidence in its own predictions, which is vital for scientific applications.