Model Expansion
- Plant species coverage expanded from 5,000 to 7,918 species
- Includes 185 tropical plant species sourced from Singapore’s National Parks Board dataset
- Model architecture updated from an EfficientNet-B4 convolutional backbone to a DINOv2 Vision Transformer fine-tuned on the PlantCLEF 2024 dataset — a collection of 7,806 plant species compiled for the international plant identification benchmark
- Weights sourced from the publicly available checkpoint published by vincent-espitalier on Hugging Face.
Accuracy
- Top-1 accuracy: 79.8% · Top-3: 92% · Top-5: 94.3% across all 7,918 species
- Existing species (5,071 classes): 81% top-1 accuracy
- Newly added species (2,847 classes): 75.7% top-1 accuracy
- When the model’s confidence score for its top prediction is 90% or above, it is correct 96.5% of the time
- For lower-confidence results, the correct species is present in the top-3 suggestions 92% of the time
Deployment
- Server-side inference optimised via ONNX Runtime, reducing model execution time by approximately 50% compared to the previous PyTorch implementation
- No action is required — the updated model has been deployed and is available to use and test in the app now on iOS and Android