Skip to content

instructions.md

Latest commit

 

History

History
39 lines (28 loc) · 2.62 KB

instructions.md

File metadata and controls

39 lines (28 loc) · 2.62 KB

Data Cleaning Strategies :contentReference[oaicite:0]{index=0}

1. CLAHE (Contrast Limited Adaptive Histogram Equalization)

  • Explanation: Standard Histogram Equalization often blows out the sky to see the ground. CLAHE breaks the image into small tiles and equalizes them individually.
  • The Logic: This ensures that if half the track is in a tunnel and half is in the sun, both segments have visible edges for the model to detect. It prevents the model from being "blinded" by harsh RC track lighting.
  • Resources: OpenCV Histograms - CLAHE.

2. Laplacian Filtering (Blur Detection)

  • Explanation: The Laplacian operator calculates the "sharpness" of an image by looking at the intensity of edges.
  • The Logic: High-speed RC cars vibrate intensely. A sharp image has high variance; a blurry one has low variance. You set a threshold and automatically delete frames that are too "mushy" for the model to learn features.
  • Resources: Blur detection with OpenCV.

3. SSIM Culling (Structural Similarity Index)

  • Explanation: SSIM looks at "perceptual" changes like texture and contrast between two consecutive frames.
  • The Logic: If your RC car is idling at the start line, you don't need 300 identical frames. SSIM identifies frames that are 95%+ identical so you can delete the extras.
  • Resources: Scikit-Image Structural Similarity.

4. PCA (Redundancy & Diversity Analysis)

  • Explanation: PCA reduces the high-dimensional pixel data (thousands of pixels) into a few "Principal Components" that represent the most significant visual variations in your dataset.
  • The Logic:
    • Clustering: If you plot your images on a 2D PCA graph, similar images (e.g., all "Left Turns") group together.
    • Balancing: If one cluster is massive (Straightaways) and another is tiny (Chicanes), you can delete frames from the massive cluster to "even out" the dataset.
  • Resources: Scikit-Learn PCA Documentation.
    https://math.stackexchange.com/questions/1520832/real-life-examples-for-eigenvalues-eigenvectors/3985012#3985012

Summary Table

Step Technique Purpose Target Metric
1 CLAHE Normalizes lighting/shadows. Balanced Color Histogram
2 Laplacian Removes motion-blurred frames. Variance Threshold > 100
3 SSIM Culling Removes adjacent near-duplicates. Similarity Score < 0.90
4 PCA Culling Ensures even distribution of track types. Even Cluster Density