plant-id

We tested LLMs to improve multilingual descriptions of plants, insects, and fungi where Wikipedia entries were inadequate. GPT-4o outperformed smaller models such as GPT-4o mini and Mistral Pixtral 12B, especially in grammar and factual accuracy. It performed better in widely spoken languages and common species. In the end, GPT-4o was used to generate descriptions in 28 languages, providing better coverage for the Kindwise API.

This feature provides clearer and more actionable insights into plant health.

The Plant.id model upgrade resulted in significant improvements in accuracy. It reached over 85% accuracy in the first response (TOP 1). The increase is greatest for European and North American plants, with 6.0% and 3.6%, respectively.

Plant.id now allows plant species to be filtered by various criteria such as geographical distribution, morphology, ecology or practical use. This increases identification accuracy and confidence in the model.

We started looking for synergies between our narrow ML models and LLMs. Let's play with this together on our demo 🙂

Multimodal Large Language models have recently made it possible to identify various objects from images, including plants. This became a competition for narrowly focused neural network models, such as that of Plant.id. Here we compare the accuracy of these models by focusing on the proportion of wrong answers.

Plant.id emerged as the leading application for the automatic identification of tree species, outperforming other methods in accuracy and efficiency.

Our endpoint allows our clients to search through huge botanical, fungi or insect database by latin name or common names to get various species info

As the world of houseplants and ornamental varieties continues to flourish, many people now seek more than species identification. Our new 'Varieties' model responds to this demand and offers you a unique opportunity to gain a competitive advantage by enabling the identification of almost 1,000 plant varieties.

We have upgraded the Plant.id model, increasing the number of classes from 33,325 to 35,756. The new version, released in October 2023, shows significant performance improvements over the previous models, especially in Europe and North America.

We boosted the number of classes from 12k to 33k and increased the accuracy!