PR Description

This PR addresses the high latency in gaze estimation by significantly reducing the SVR hyperparameter search space through systematic optimization.

Fixes : #65

Problem Discovery

Initial testing revealed that the SVR model was testing 75 parameter combinations (5×5×3), causing processing times of 4+ minutes for 80 samples. This was identified as the primary bottleneck in gaze estimation performance.

Investigation Process

We conducted multiple experiments to find the optimal balance between speed and accuracy:

1. Kernel Reduction Experiment: Initially tried [linear] only → Achieved 3.5s but sacrificed accuracy potential
2. Balanced Approach: Kept all kernels but reduced C and gamma ranges → Achieved 2.73s with better accuracy preservation
3. Scalability Testing: Verified performance with varying sample sizes to ensure consistent results (80 samples & 560 samples) from calibration step

Final Solution

• Reduced C values from [0.1, 1, 10, 100, 1000] to [1, 10] (5 → 2 options)
• Reduced gamma values from [0.0001, 0.001, 0.01, 0.1, 1] to [0.01, 0.1] (5 → 2 options)
• Kept all kernel types ["linear", "rbf", "poly"] to preserve accuracy potential (3 options maintained)

Results

• Total combinations reduced from 75 to 12 (84% reduction)
• Processing time improved from 4+ minutes to 2.73 seconds for 560 samples
• Maintains accuracy potential by preserving multiple kernel types
• Scalable solution that maintains performance across different sample sizes