Reproduce “FruitEnsemble: MLLM-Guided Arbitration for Heterogeneous ensemble in Fine-Grained Fruit Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Enhui Yu, Junhui Li, Ruitong Lu, Jialu Li, Youshan Zhang
FruitEnsemble, developed by researchers from University of Science and Technology Liaoning, introduces a dynamic framework for fine-grained fruit recognition, combining a heterogeneous visual ensemble with selective Multimodal Large Language Model (MLLM) arbitration. The system achieves 70.49% Top-1 accuracy and 91.50% Top-5 accuracy on the new Fruit-306 dataset, maintaining an average inference latency of 19.8 ms by invoking the MLLM for only 15% of samples.
Key insights
- Dynamically routing only genuinely ambiguous cases to an MLLM, guided by visual ensemble uncertainty and domain-specific knowledge, effectively leverages MLLM reasoning power without incurring prohibitive computational costs for all samples.
- The creation of a comprehensive, multimodal dataset with expert-curated textual botanical descriptions is essential for training and validating hybrid vision-language systems in fine-grained recognition tasks.
- A heterogeneous ensemble, enhanced by uncertainty-aware dynamic aggregation and hard sample-aware joint optimization, significantly improves discriminative power and robustness for visually similar categories and long-tail distributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.