Train an AI4EOSC module (using YOLO)
Train an AI4EOSC module (using YOLO). - Dr. Enoc Martínez (UPC)
OBSEA SEafloor Observatory
- Underwater cabled observatory.
- Located at NW Mediterranean Sea.
- Shallow waters (20 m depth).
- Multiparametric Observatory.
Ecosystem Monitoring
- Multiple cameras deployed.
- Image archive since 2011.
- Fish abundance & behavior studies.
- Image analysis - time consuming!
Traditional Ecosystem Monitoring
- Very time consuming.
- Not reproducible.
- Expertise in biology.
- Extremely repetitive task.
Doing this for tens of thousands images for hundreds of different species is incredibly time consuming!
AI-based Ecosystem Monitoring
- IT skills required.
- Significant effort to set up, but…
- Fully automated workflow.
- No human intervention required.
Let scientists do science, not count fish!
iMagine Project
- AI services for marine science.
- Focuses on image processing.
- Access to AI platform.
- More info at: https://imagine-ai.eu
Object Detection in a Nutshell
1. Collect your pictures.
2. Label your data.
3. Select a model.
4. Format your data.
5. Train the model with your data.
6. Use your model (inference).
Object Detection models
- Mature field, rapidly evolving.
- Lots of open-source tools.
- State-of-art model reviews.
- https://paperswithcode.com
- Our selection: YOLOv8
- Good balance between precision and speed.
- Easy-to-use.
- Good documentation.
Training a YOLOv8 module
- Training parameters:
- epochs
- image size
- batch size
- Performance metrics:
- precision/recall
- intersection over union (IoU)
- mean average precision (mAP)
mAP@0.5
Some results
AI Platform
- Dedicated resources for training.
- Off-the-shelf AI modules.
- Interaction with CLI and/or API.
- Open-source, containerized:
- github
- docker hub
- More info at https://www.imagine-ai.eu/
Training a YOLOv8 module (jupyter)
This part of the course is very practical so most of the information comes from watching the video.
- Deploy a YOLOv8 module in the AI platform.
- Platform https://dashboard.cloud.imagine-ai.eu/marketplace
- Download the dataset. https://universe.roboflow.com/obsea/test-yolov8-ai4eosc/dataset/1
- Train with custom data with jupyter cli.
- YOLOv8 docs https://github.com/ai4os-hub/ai4os-yolov8-torch
- Import trained model
- OBSEA fish detector YOLOv8 model https://dashboard.cloud.imagine-ai.eu/marketplace/modules/uc-enocmartinez-deep-oc-obsea-fish-detection
Conclusions
- User friendly and easy-to-use platform.
- Pre-trained models in the marketplace.
- on-demand resources for training/inference
- command-line / API interfaces