Q1 2025 Roboflow Update

We aim to build a model that produces results similar to the table shown in the reference paper (below), for coral restoration reefs and/or natural reefs.

 

To date, the following are the model evaluations:

• The Mean Average Precision (mAP)/CT is still low

• Unbalance trained images between lifeforms (actually, this is how nature works. some lifeforms dominant, and some not)

• Good enough to detect common/dominant coral lifeforms but with CT around 5-50%

• Hard to detect small objects. i.e. Acropora fragments that placed just above the metal frame

• Confusion to detect similar lifeforms: (AC branching & C branching), (C submassive & AC Digitate), (C foliose, AC tubulate, Sponge with foliose form)

What’s next?

Current dataset:

• Dataset Recomposition of Train, Valid, Test images.

• Try to deploy a model with the result in table/graph forms.

Next dataset:

• Annotate more images. 1. Images in metal frames, frames, or naturals in more diverse categories. 2. Try to train more images that previously have small numbers. 3. Better to not use photogrammetry images (because try to train detecting small objects first, need clear pictures).