A segmentation project done as the final project of the Machine Learning for Computer Vision course at ESILV
A detailed presentation can be found in this repo at Project_presentation.pdf
- Definition: Label each pixel of an image with a class/category.
- Objective: Segment underwater images into objects (e.g., fish, reefs, plants).
- Goal: Automatically label each pixel for underwater imagery analysis.
- Challenges in segmenting underwater imagery.
- Importance of precise segmentation in marine applications.
- Train a segmentation model to identify underwater objects.
- Generate segmentation masks with pixel-wise classification.
- Train/Validation: 1525 paired samples.
- Test: 110 paired samples for benchmarking.
| Category | RGB Code |
|---|---|
| Fish | Yellow: 110 |
| Reefs | Pink: 101 |
| Aquatic Plants | Green: 010 |
| Wrecks/Ruins | Sky: 011 |
| Human Divers | Blue: 001 |
| Robots | Red: 100 |
| Sea-floor | White: 111 |
| Background | Black: 000 |
- Model: U-Net architecture.
- Enhancements: Dropout layers to prevent overfitting.
- Techniques: Horizontal flipping, rotation, Gaussian noise, etc.
- Dice Loss: Measures overlap between predicted and ground truth masks.
- IoU: Evaluates the intersection over union for object regions.
- Adam Optimizer: Learning rate = 1e-4.
- Scheduler: Reduces learning rate when validation loss plateaus.
- Steady decline over 50 epochs.
- Minimal gap between training and validation loss.
- Significant improvement in training and validation IoU.
- Hyperparameter Tuning: Optimize learning rates and batch sizes.
- Dataset Expansion: Incorporate similar datasets.
- Transfer Learning: Use pretrained models for better feature extraction.
- Advanced Augmentation: Focus on underrepresented classes.
- Marine Ecology: Track underwater species.
- Archaeology: Identify submerged historical sites.
- AUVs: Enhance underwater navigation.
- Environmental Monitoring: Assess vegetation health and pollution.