Color Constancy and Illuminant Estimation

Overview

A. Understanding the Color Constancy Problem

1. Illuminant Estimation for macbeth_syl Image

• Selected a white patch and computed the average RGB value: [238, 220, 250]
• Normalized to max value 250 using a scaling factor

2. Illuminant Estimation for macbeth_solux Image

• Repeated the same process
• Resulting RGB value: [132, 159, 250]

3. Angular Error Between Light Sources

• Computed angular error using normalized dot product:
  • Result: 13.84°

4. Image Remapping Under Canonical Light

• Used diagonal model to map original blueish image to canonical lighting
• Compared original, corrected, and canonical images
• Observed that corrected image visually approximates the canonical, with small discrepancies in brightness

5. RMS Error of Chromaticity Coordinates

• Filtered out dark pixels (R+G+B < 10)
• Computed chromaticity (r, g) and Euclidean distance for RMS error
• Results:
  • Original vs Canonical: 0.1149
  • Corrected vs Canonical: 0.0475

6. MaxRGB Algorithm for Remaining Scenes

• Estimated light color using MaxRGB:
  • apples2: [182, 228, 250], angular error: 9.80°
  • ball: [153, 177, 250], angular error: 3.57°
  • blocks1: [250, 242, 209], angular error: 19.95°
• Manually excluded artifact regions to improve estimate in ball scene

7. Final Image Triplet Display and RMS Report

• Generated corrected images and computed RMS error for each:
  • apples: 0.0608
  • ball: 0.0494
  • blocks: 0.0984
• RMS error rankings match angular error rankings, confirming consistency

8. Comparison with Gray-World Method

• Computed gray-world estimates and angular errors:
  • apples: [49, 78, 250], error: 19.62°
  • ball: [129, 167, 250], error: 1.39°
  • blocks: [250, 154, 180], error: 23.43°
• RMS errors:
  • apples: 0.1382
  • ball: 0.0554
  • blocks: 0.1514
• MaxRGB outperformed gray-world overall, despite an anomaly in ball scene where gray-world had lower angular error but higher RMS

---

Conclusion

This assignment implemented two methods of light estimation (MaxRGB and Gray-World), applied them to real images using the diagonal model, and evaluated their accuracy using angular and RMS error. The results support MaxRGB as a more consistent performer on this dataset, though scene-specific anomalies reveal the complexity of color constancy in practice.