Sayan Ghosh, Michael Rakowiecki, Atishay Singh, Hao Wang, Michael Wolf
Object Detection for aquatic environments based on Detectron2.
This project performs object detection on aquatic images using Faster RCNN R101 from the Detectron2 Model Zoo as a backbone. Our model was trained on the MODD2 dataset using this pre-trained model. We used data augmentation to increase the size of the training set and to make the model more robust.
See our paper to learn more about the project.
We use the MODD2 dataset created by Bovcon, Borja and Muhovi, Jon and Per, Janez and Kristan, Matej.
This project was created to train a model on a dataset that contained a largely monotonous foreground with weak distinguishing features for the objects. This is the first step in our plan to create a model to perform obstacle detection for an autonomous robot (see MRover) in the desert.
Use our Colab Notebook and follow the instructions to download the dataset, train the model, and evaluate images.