Supporting Ultra-High-Resolution Digital Agriculture Tasks with Fully Synthetic Curriculum Learning

Jacob Hatef, Quentin Anthony, Nawras Alnaasan, Dhabaleswar Panda; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2026, pp. 491-499

Due to the quadratic cost of attention with image size, foundational models in computer vision are pretrained on a large selection of low-resolution images, such as ImageNet-21k, ImageNet-1k, LAION-400M, etc. These models are then finetuned for downstream tasks using domain adaptation. However, for high-resolution tasks, the model must learn to adapt to two different axes: the dataset distribution and the image resolution. Borrowing from Natural Language Processing, we propose a multi-staged curriculum learning-based strategy for adapting pretrained models to high-resolution. We use ImageNet-1K and super-resolution models to create a high-resolution, multi-staged, synthetic dataset that matches the pretraining data distribution of ViT-L/16. After our curriculum learning step, we finetune on two downstream, high-resolution, digital agriculture classification tasks. Our proposed high-resolution adaptation step creates high-quality models that outperform the baseline pretrained model by 1.40% to 2.48%. Our technique is orthogonal to architecture choice and can be applied to any pretrained model.