尺寸至关重要：从单目图像重建真实尺度三维模型用于食物份量估计

The increasing prevalence of chronic diseases linked to diet, such as obesity and diabetes, underscores the critical need for accurate food intake monitoring. While AI-driven dietary assessment has advanced significantly in recent years, the ill-posed problem of recovering size (portion) information from monocular images for precise estimation of “how much was consumed?” remains a pressing challenge. Existing 3D reconstruction methods have achieved impressive geometric fidelity but often fall short in addressing real-world scale recovery. This research aims to overcome this limitation by developing a novel approach capable of reconstructing real-scale 3D food models from monocular images. The core idea involves integrating deep learning with geometric constraints to resolve the inherent scale ambiguity in monocular vision. Specifically, a multi-task learning framework is introduced, comprising a depth estimation network for predicting depth maps and a size regression network for predicting object size references. These depth maps and size references are subsequently integrated into an optimization process to generate 3D mesh models with accurate physical dimensions. To enhance reconstruction accuracy, prior knowledge-based food shape constraints and textural information are leveraged. Extensive experiments on a diverse dataset of food images demonstrate the superior performance of this method in recovering the true dimensions and shapes of food items, thereby laying the groundwork for more accurate food portion estimation. Compared to methods focusing solely on geometric reconstruction, this approach exhibits significant improvements in metric scale, effectively addressing the difficulty of obtaining actual size information from monocular images.