Today, we're sharing the story of our venture to create a rock and mineral identification API. We set out to create a product that would be able to identify the most common rocks and minerals. The journey, however, will be much longer than expected.
The first version of the model falls short of our expectations, so we've decided to postpone the project and focus our efforts elsewhere (such as improving our favorite beta products - Insect.id and Mushroom.id API).
If you're still interested in trying out our model as it is, let us know at firstname.lastname@example.org we'll give you access to our internal demo.
How we tried to find the gold:
1. Working with the experts
It is important to us that our products contain the expertise from the chosen field. That's why we worked with geology experts who helped us formulate the problem and create the list of classes we wanted to include in the model.
2. Data collection
Obtaining a comprehensive dataset of mineral and rock images was a significant hurdle. The variety and quality of images available were limited, making it difficult for our model to learn effectively. There is also a degree of variability in how a rock can look, as well as similarity between different rocks. For the future, we are considering expanding our database by having more images annotated by our experts.
3. Model training
We used our extensive experience in developing machine-learning models for image recognition. Unfortunately, the accuracy and performance we achieved with the available data was not satisfactory. The model with 50 classes of rocks and minerals correctly labelled 45% of the images (when looking at the TOP1 label). We were hoping for a much higher level of accuracy to ensure a stable quality of our products.
4. I iterate, therefore I am
We tried many different approaches to improve the model (such as adding more images to the training dataset and making prioritisation of the most common rocks). Eventually, we decided that our efforts were not bearing much fruit and the project was paused.
We keep moving forward
Our attempt to create an API for rock and mineral identification may not have resulted in the reliable and accurate product we had hoped for, but it has been a journey of learning and resilience.
We have decided to put the project aside for now, but it is ready to be picked up and improved upon at any time. So if you have any ideas on improving our API, want to collaborate on future projects, or just want to test the API as it is, we would love to hear from you. Feel free to contact us at email@example.com.
The new API can identify 3,100 species of fungi (including lichens and related organisms such as slime molds) and includes enhanced information that has been carefully selected and prepared based on the needs of people interested in mushroom identification.
Plant.id is an API that identifies plant species and diseases from photos with machine learning. Send us images of your plant and get the possible suggestions with plenty of other information including representative images of the species.
The primary purpose of Plant.id is to provide a plant species and diseases identification API. To make it accessible to non-programmers, we created a web demo, which you can try at https://plant.id/ website.
Let us introduce you to the first production version of Plant.id Health Assessment. During the last year, we have been collecting feedback on the beta version of our plant disease identification API. Moreover, we have expanded our datasets and deepened our understanding of machine learning-powered plant disease identification.