BRE04: "Voice in the Blue" Real-time speech annotation of diver-recorded underwater video.

We know more about the surface of the Moon than we do about the mapping of the deep seafloor, but artificial intelligence combined with autonomous underwater robots may soon turn the tide. Indeed, exploration vehicles are becoming increasingly autonomous, allowing us to explore our oceans, discover marine fauna and flora, and better understand our blue planet. However, although artificial intelligence can automate image recognition tasks, these techniques require datasets to learn how to better recognize the environments in which they are deployed. Unfortunately, underwater databases are very limited and suffer from a lack of annotations - that is, descriptions of what is visible in each image. Therefore, in this challenge, we propose to rely on the diving community to annotate their dive images in real time. To achieve this, we envision equipping divers with GoPro-type cameras and a voice capture system to produce new annotated datasets. Through this citizen science approach, we hope to train robust new AI models for deep-sea exploration. This project is part of the ANR CESAR initiative.

The goal is to develop an automatic annotation algorithm for underwater data. We envision the following steps in the challenge:
1. Handling the video data
2. Creating audio data associated with the videos (to simulate the audio that the diver might produce in real time during a dive)
3. Proposal of an AI model to transcribe the audio annotations into images (segmentation, for example)
4. Evaluation and prospects for the project

AI Developer
Data Analyst
Seabed experts