Overview / Usage

Coral species detection underwater is a challenging problem. There are even cases when the experts (marine biologists) fail to recognize corals, hence limiting ground truth annotation for training a robust detection system. Identifying coral species is fundamental for enabling monitoring of coral reefs, a task currently performed by humans, which can be automated with the use of underwater robots. By employing temporal cues using a tracker on a high confidence prediction by a traditional object detector, we can augment the collected dataset for the retraining of the object detector. However, using trackers to extract hard examples, recklessly is counter-productive and will deteriorate performance. In this work, we show that using a simple deep neural network for image quality can help regulate hard sample extraction. We empirically evaluate our approach in a coral object dataset, collected via an Autonomous Underwater Vehicle (AUV) and human divers, that shows the benefit of incorporating hard examples obtained from tracking. This work also demonstrates how controlling sample generation by tracking using a simple deep network can further improve an object detector.

Methodology / Approach

The project utilizes a coral detection and tracking system already developed in an earlier project. Using the detection and tracking system, the approach generates a tremendous amount of pseudo-labeled data. Later, this pseudo-labeled data is used in the training processin a carefully constrained manner.