The precise recognition of food categories is crucial for health management. Existing datasets like Food-101 and VIREO Food-172 are well-curated but lack representation of daily-life scenarios. To bridge this gap, two new benchmarks, DailyFood-172 and DailyFood-16, have been introduced. The MCRL method is proposed to handle the challenges posed by the variance in food appearances between curated datasets and real-life scenarios. By dynamically aligning target samples with multiple source clusters, MCRL enhances classification accuracy and generalization abilities.
The study highlights the discrepancy in appearance consistency between dishes from curated datasets and those from daily meals. It emphasizes the need for datasets that better represent everyday food images for practical applications. The proposed MCRL method addresses the "category ambiguity" problem by dynamically learning distribution shifts towards multiple source cluster features during training.
Through experiments on extensive visual cross-domain tasks, it is demonstrated that integrating MCRL with conventional UDA methods significantly improves classification accuracy in target domains. The ablation study further validates the effectiveness of MCRL in dynamic distribution alignment between domains. Qualitative results showcase instances where MCRL outperforms state-of-the-art methods in accurate food classification.
Til et andet sprog
fra kildeindhold
arxiv.org
Vigtigste indsigter udtrukket fra
by Guoshan Liu,... kl. arxiv.org 03-13-2024
https://arxiv.org/pdf/2403.07403.pdfDybere Forespørgsler