AutogradeInnovation_AIC

Competition Information

For more details about the competition, please refer to this notebook.

Idea

Details of the core idea can be found in idea.md.
Additionally, the inspiration for this idea is documented in this notebook.

Progress Updates

• 16/12/2024

  • Received the dataset and made initial observations on both the data and the publicly available solution.

• 17/12/2024

  • Created a simple web-based image reader to identify labeling errors.
  • Formulated assumptions about these errors and validated them.

• 18/12/2004

  • Removed duplicate labels.

• 19/12/2024

  • Conducted a basic YOLOv8 fine-tuning.
  • Began contour detection.

• 20/12/2024

  • Continued contour detection work.

• 21/12/2024

  • Completed the contour detection process.

• 22/12/2024

  • Began the first testing phase.
  • Performed image cropping.

• 23/12/2024

  • Further removal of duplicate labels.
  • Removed redundant labels.

• 24/12/2024

  • Addressed missing labels.
  • Replaced bounding boxes that had two labels with the correct label.
  • Cropped images for Test Set 1.
  • Fine-tuned YOLOv8 on these cropped images.
  • Ran inference on Test Set 1 with the fine-tuned model.

• 25/12/2004

  • Downloaded images for Test Set 2.

My Solution

Training Phase

Note: All the progress is also documented in data_analysis.ipynb.

1. Label Error Detection
   Used a simple web-based image reader (found in image_reader) to identify labeling errors.

2. Label Error Correction
   Cleaned the labels by removing duplicates, redundancies, and missing entries, and by replacing bounding boxes that had multiple labels. All related work is in the data_engineering folder.

3. Model Training
   Trained a YOLOv8 model on Kaggle. You can find the configurations and training details in model_selection.ipynb.

Testing Phase

1. Contour Detection & Cropping
   Used the previously defined contour detection function to crop images.
   For images where contours were not automatically detected, contours were drawn manually.

2. Inference
   Cropped images were then passed through the fine-tuned model for final inference.

File Structure

• data_analysis.ipynb
  Contains data exploration and ongoing progress notes.

• image_reader
  A simple web-based viewer that displays images and their labels.

• data_engineering
  Includes label error corrections and image processing workflows:

  • contours_detecting.ipynb: Detects image contours.
  • contours_sorting.ipynb: Sorts detected contours.
  • image_cutting.ipynb: Crops images based on identified contours.
  • label_processing.ipynb: Handles label errors (duplicates, missing entries, redundancies).
  • label_correting.ipynb: Fixes bounding boxes with multiple labels.

• model_selection.ipynb
  Documents the model choice for the competition along with training configurations and processes.