Multi-Scale ViT Inference with Habitat-Fit Priors and kNN Retrieval for Multi-Species Plant Identification

arXiv cs.AI · 6 hours ago

DS@GT ARC placed third in the PlantCLEF 2026 plant identification challenge using a pipeline built on fine-tuned DINOv2 ViT-L/14 applied to multi-scale image tiles with kNN retrieval and habitat-aware post-processing. The selected submission achieved a macro-F1 score of 0.43902 on the private leaderboard, with habitat-fit demotion and multi-scale aggregation identified as the largest performance contributors in ablation studies. The approach demonstrates that geographic and altitude priors combined with multi-scale inference can improve plant species detection in high-resolution vegetation plot images despite training only on single-plant labeled data.