Zero-Shot Segmentation for Multi-Label Plant Species Identification via Prototype-Guidance
Published:Dec 24, 2025 05:00
•1 min read
•ArXiv AI
Analysis
This paper introduces a novel approach to multi-label plant species identification using zero-shot segmentation. The method leverages class prototypes derived from the training dataset to guide a segmentation Vision Transformer (ViT) on test images. By employing K-Means clustering to create prototypes and a customized ViT architecture pre-trained on individual species classification, the model effectively adapts from multi-class to multi-label classification. The approach demonstrates promising results, achieving fifth place in the PlantCLEF 2025 challenge. The small performance gap compared to the top submission suggests potential for further improvement and highlights the effectiveness of prototype-guided segmentation in addressing complex image analysis tasks. The use of DinoV2 for pre-training is also a notable aspect of the methodology.
Key Takeaways
Reference
“Our solution focused on employing class prototypes obtained from the training dataset as a proxy guidance for training a segmentation Vision Transformer (ViT) on the test set images.”