[D] Working on a photo-based calorie tracker app

An independent developer is tackling one of computer vision's most persistent real-world challenges: building a truly accurate photo-based calorie tracker. While apps like CalAI exist, they often falter with complex mixed dishes and unreliable portion size estimation. This new project aims to move beyond treating the problem as a simple API call to a general vision model like OpenAI's GPT-4V. Instead, the developer is taking a serious ML-first approach, experimenting with specialized architectures like YOLOv8 for multi-food object detection and actively researching whether to add segmentation or regression models specifically for volume and portion estimation.

The technical core of the challenge lies in moving from simple food identification to precise 3D understanding from a 2D image. The developer is seeking community input on key architectural decisions: whether a detection-plus-regression pipeline or a full segmentation approach is superior, the availability of portion-aware food datasets, and the practicality of implementing monocular depth estimation on mobile devices. Success would mean creating a tool that doesn't just label a 'burrito bowl' but can estimate the volume of rice, beans, and salsa separately to calculate a precise calorie count, representing a significant leap in applied AI for health and nutrition.