Skip to content

Latest commit

 

History

History
89 lines (64 loc) · 4.4 KB

PROCESS.md

File metadata and controls

89 lines (64 loc) · 4.4 KB
Raw

Development Process: Rescuer-seeker

1. Project Initialization and Data Preparation

  • Started with a dataset of aerial imagery from disaster scenarios(https://universe.roboflow.com/deep-learning-dohjx/natural-disater). DISCLAIMER: Dataset utilizes black and white images, when tested with regular colored images and video it showed better performance
  • Implemented data preprocessing pipeline, including resizing to 640x640 pixels and augmentation techniques.

2. Model Selection and Configuration

  • Chose YOLOv11m as the base model for its balance of speed and accuracy.
  • Configured the model for three classes: Houses, People, and Vehicles.
  • Initialized with pretrained weights and adapted the model architecture for the specific task.

3. Training Process and Optimization

  • Implemented a training pipeline using Ultralytics framework.
  • Utilized a NVIDIA GeForce RTX 2070 Super GPU for training.
  • Key training parameters:
    • Batch size: 8
    • Initial learning rate: 0.0005
    • Optimizer: SGD with momentum (0.937) and weight decay (0.0005)
  • Implemented learning rate scheduling with cosine annealing.
  • Applied various data augmentation techniques including mosaic, mixup, and copy-paste.

4. Challenges and Solutions

  1. Python Version Compatibility:

    • Challenge: Persistent warnings about Python version (3.9.13 installed, >=3.10 required).
    • Solution: While not resolved during this training run, future iterations should update the Python environment to 3.10 or higher.

  2. Disk Space Management:

    • Challenge: Warnings about insufficient disk space for caching images.
    • Solution: Adjusted caching strategy to work within available disk space constraints.

  3. Dataset Inconsistencies:

    • Challenge: Mismatch between box and segment counts in the dataset.
    • Solution: Focused on box detection, removing segments to ensure consistency, making it about what it's about, finding people and saving lives.

  4. Resource Optimization:

    • Challenge: Balancing model performance with computational resources.
    • Solution: Carefully tuned batch size and image size to maximize GPU utilization without exceeding memory limits.

5. Training Progression

  • Initiated training for 200 epochs, but stopped at 130 epochs due to satisfactory performance.
  • Observed consistent improvement in model metrics throughout training:
    • mAP50 increased from 0.263 in epoch 1 to 0.695 by epoch 130.
    • Precision improved from 0.325 to 0.729.
    • Recall enhanced from 0.349 to 0.668.

6. Evaluation and Testing

  • Implemented comprehensive evaluation on the test set (522 images).
  • Achieved impressive final metrics:
    • mAP50: 0.702
    • mAP50-95: 0.459
    • Precision: 0.751
    • Recall: 0.586

7. Performance Analysis

  • Conducted per-class performance analysis:
    • Houses showed high precision (0.776) but lower recall (0.536).
    • People detection was the most accurate with high mAP50 (0.784) and balanced precision and recall.
    • Vehicles had the lowest precision (0.662) but decent recall (0.554).

8. Iterative Improvement

  • Regularly monitored training progress and adjusted parameters as needed.
  • Implemented early stopping at epoch 130 due to satisfactory performance and to prevent overfitting.

9. Documentation and Repo Organization

  • Created comprehensive README.md, MODEL.md, and PROCESS.md files.
  • Organized repository structure for clarity and reproducibility.
  • Documented code with clear comments for maintainability.

Conclusion

The development of Rescuer-seeker demonstrates the successful adaptation of state-of-the-art object detection technology to the critical domain of disaster response. Through careful model selection, dataset preparation, and iterative optimization, we've created a tool with the potential to significantly enhance the efficiency and effectiveness of aerial search and rescue operations.

Key achievements:

  • Successful training of a YOLOv11m model on a custom disaster scenario dataset.
  • Achieving high accuracy in detecting crucial elements (houses, people, vehicles) in aerial imagery.
  • Overcoming technical challenges related to data processing and resource management.
  • Creating a model that balances accuracy with the speed necessary for real-time applications in critical situations.

This project showcases the potential of AI to make a meaningful impact in disaster response and humanitarian efforts, paving the way for more advanced and efficient search and rescue technologies.