Heng Song, Junhao Zhu, Tong Liu, Xinpeng Lu, Junwu Zhu · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3835535
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Estimating dish calories solely from visual data is an ill-posed problem due to the absence of scale references and hidden ingredients. While Multimodal Large Language Models (MLLMs) offer rich semantic understanding, they are prone to hallucinations and lack the reliability required for safety-critical dietary assessment. To address this, we introduce MARC (Multimodal Anchored Residual Correction), a framework that reformulates calorie estimation as a conditional residual prediction task. Rather than predicting absolute values directly, MARC grounds its inference on a retrieval-based USDA anchor. Specifically, we model the refinement process as a contextual routing problem where a policy network adaptively selects the most effective strategy, ranging from visual scaling to compositional rectification, to predict the precise residual term. To enhance reliability, we incorporate a lightweight gated calibration mechanism to mitigate overconfidence and employ conformal prediction for rigorous uncertainty quantification. Furthermore, we release the CCM, a large-scale Chinese Cuisine Multimodal dataset containing over 32,000 dishes, including an expert-verified gold-standard subset. Extensive experiments on CCM and the public Nutrition2.7K dataset demonstrate that MARC achieves state-of-the-art performance, significantly outperforming existing visual regression models and general-purpose MLLMs while offering valid confidence intervals for trustworthy dietary assessment. Our code is available at https://github.com/Agentyzu/MARC.
ACM International Conference on Multimedia (ACM MM) 2026 · 0 citations
No comments yet — start the discussion below.