How Dan Trained RF-DETR

We want to create an AI detector capable of detecting any fishy-looking creature anywhere. For AI for Conservation folks, that would be a MegaDetector for Underwater fish species.

We know that the amount of data publicly available online, could support such effort. Therefore, we need to collect all that data, and compile it into a single database, which we can use to train the model.

This project is a community effort. We have a lot of datasets to process, but if each one of us helps processing a few of them, we can create the most comprehensive fish database and generate an open-source model that can greatly benefit the community.

Dataset Creation Process

There are many useful fish datasets online. We need to find them and process them in a common format, to generate a comprehensive fish dataset.

We use the following excel to keep track of such work.

https://docs.google.com/spreadsheets/d/1hd24tHF1cYDses96-iPREuf0NR9mOGHUb0TKok4ME7M/edit?gid=0#gid=0

For each dataset, we follow the steps below:

  1. Find the Dataset: we read a paper which shares datasets of fish with bounding boxes, and save it to the dataset file.
  2. Understand the Dataset: based on the information shared in the paper/website, we understand the location, species and other metadata about the data
  3. Analyze the Data: we download the dataset and perform data analysis to calculate the dataset size, number of annotations etc.
  4. Process the Data: we generate reusable code to download, process and visualize the dataset. So that others can benefit from it.

🚀 How can I contribute?

We need help writing code to process each dataset listed in the excel file above.

The goal is to have for each dataset a notebook which downloads the data and visualizes random instances with their annotations.

If you want to support this project, please pick a dataset and develop code to process it.

To do so, follow these steps:

  1. Choose a dataset from the excel and add your name to the column Data Processing Owner (DPO) and your email to DPO's email
  2. Analyze the dataset, and fill in all the columns of Step 2 and Step 3 in the excel.