| Figure A - Dolphin in the Wild | Figure B - Fin Detection Result |
|---|---|
 |
 |
| Original field image before model inference. | Model output with dolphin fin localized by a bounding box. |
Legend
Figure A: raw dolphin image collected in field conditions.Figure B: automatic dolphin fin detection used before human review.
A computer vision project designed to support dolphin monitoring and conservation from photo datasets. The work is organized in two stages:
- Stage 1: detect dolphin fins in images.
- Stage 2: identify individual dolphins from detected fins.
| Item | Details |
|---|---|
| Domain | Marine computer vision |
| Primary Goal | Build an end-to-end dolphin monitoring workflow from raw photos |
| Stage 1 | Fin detection (completed) |
| Stage 2 | Individual identification (in progress) |
| Data Source | Dolphin image dataset prepared from field collections and Roboflow |
| Human-in-the-loop | Yes, via review workflow for prediction validation |
Field campaigns can generate thousands of dolphin images. Manual review alone is slow and difficult to scale.
This project helps teams move faster by pre-detecting fins, organizing candidate images, and preparing data for expert validation. The final objective is to make long-term dolphin tracking more reliable by combining model assistance with human expertise.
- Built and validated a dolphin fin detection stage with strong results.
- Generated visual outputs and prediction artifacts for review.
- Set up the foundation for an active-learning style loop with human validation.
The active phase is individual dolphin identification. The goal is to connect each detection to a specific dolphin identity over time, enabling stronger ecological monitoring and research insights.