Asking Gemini to find vcuts

Asking Gemini to write code to find vcuts¶

Gemini continues to blow my mind.

I submitted the following prompt to Gemini Pro 3.1.

I am developing python code to find damage to coconut palms caused by coconut rhinoceros beetles which cause distinctive v-shaped cuts to fronds. These can be modeled as convexity hull defects in the crowns of coconut palms. I am starting with a binary segmentation mask for a badly damaged coconut palm. I want to follow these steps: 1. Cut the mask into two parts: crown and trunk. 2. Enclose the crown in a convex hull and find convexity defects. 3. Visualize results. Please provide two python code examples which perform these steps. Use opencv for the first example and scikit-image for the second example.

Here is the response from Gemini with the code it wrote for me transcribed into a Jupyter notebook.

And here’s an updated version of the same Jupyter notebook after a few fixes and improvements.

Results¶

Gemini did an amazing job considering the complexity of my prompt. Two python code examples for finding and visualizing v-shaped cuts in coconut palms were returned as requested. Gemini also returned a function which generates a synthetic binary mask of a palm with V-shaped cuts. Generation of synthetic data was not mentioned in my prompt. Both code examples executed without modification.

Comments on code for Method 1: Using OpenCV¶

When I ran the code “out of the box”, the figure showed an obvious problem which is easy to fix. The simulated v-shaped cuts are too deep, cutting the crown into two segments.

The code uses the standard convexity hull defects algorithm.

Comments on code for Method 2: Using Scikit-Image¶

Works as expected without any changes.

The code does not use the standard convexity hull defects algorithm. Instead, it generates a convex hull image and subtracts the original mask from it. The remaining pixels represent the “empty space” or defects. This is a much simpler algorithm and a lot more general. The defects are no longer constrained to being triangles as they are in the first method.

General comments¶

Gemini and other AIs are great learning tools, but only if my prompts ask the right questions.

Gemini suggested a much better approach for detecting v-shaped cuts than the standard convexity hull defects method I have been using. In addition, it looks like scikit-image might be a better python package than opencv for geometric morphometrics.