A Large-Scale Study on the Accuracy vs Cost Trade-offs of Training and Evaluation Settings in Fine-Grained Image Recognition

A large fine-grained image recognition study ran 2,000+ experiments to map accuracy-cost trade-offs across backbones, datasets, and training settings.

It found that data-aware training choices, including a new cross-image mixing augmentation for CAL, can preserve strong accuracy while reducing compute.

The authors also proposed an evaluation-only shortcut that avoids expensive discriminative crop inference yet stays competitive in accuracy.

Overall, the work helps identify the most compute-efficient recipes for fine-grained vision without giving up much performance.