Title: PerBite: A Curated Diagnostic Workflow for Bite-Aware Food Volume Estimation

Abstract:

Is it possible to rely on visually realistic food meshes for accurately estimating the volume of ingested food? This study, referred to as \method, addresses this inquiry by utilizing specific paired states—prior to and following consumption—drawn from the MetaFood CVPR 2026 Continuous 3D Reconstruction While Eating Challenge dataset. The proposed workflow employs a meticulously curated reconstruction pipeline: SAM~3 is utilized to segment regions corresponding to both the food and the plate. Subsequently, Hunyuan3D/SAM~3 produces a dimensionless mesh of the food item. To establish metric scale, the diameter of the plate is measured. Following this, Blender is used to eliminate the plate geometry, after which the remaining mesh undergoes hole-filling and watertight processing to facilitate volume integration. While MoGe-2 serves as a secondary aid for estimating dish diameter when direct measurement is ambiguous, it does not function as the primary scaling source for the challenge results presented here.

The \method approach secured first place, achieving an average Chamfer distance of 8.31 across 34 meshes, calculated via rigid ICP without scale correction. When analyzing 17 before-and-after pairs, the system attained a state-level volume Mean Absolute Percentage Error (MAPE) of 33.87%, with no monotonicity violations observed. However, the MAPE for consumed volume was recorded at 53.74%. These findings suggest that for effective dietary assessment, critical components such as surface reconstruction, metric scaling, controlled mesh cleanup, watertight volume integration, and physical depletion consistency should be evaluated as distinct entities. Source code and evaluation scripts will be hosted at \href{https://github.com/GCVCG/PerBite-CVPR-MetaFood-2026}{github.com/GCVCG/PerBite-CVPR-MetaFood-2026}.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC