High Latency in Gaze Estimation Due to Excessive SVR Hyperparameter Combinations

[sohampirale]

Issue Description

The Support Vector Regressor (SVR) model in the eye tracker API currently tests 75 different hyperparameter combinations during gaze estimation:

• 5 C values: [0.1, 1, 10, 100, 1000]
• 5 gamma values: [0.0001, 0.001, 0.01, 0.1, 1]
• 3 kernel types: ["linear", "rbf", "poly"]

This results in extremely slow processing times (over 4+ minutes for 80 samples !! 12 GB RAM + SSD), resulting in huge latency.

---

Root Cause Analysis

During investigation, I discovered that the system was testing every possible combination of hyperparameters (5×5×3 = 75), which is computationally expensive for gaze estimation. The excessive parameter combinations were primarily affecting the Support Vector Regressor model.

---

Performance Impact

• Poor user experience due to long delays
• Makes real-time eye tracking impractical
• Affects overall application responsiveness

---

Trying different hyperparamet configs

I tested parameter reductions to find the optimal balance between speed and accuracy:

1. First reduced kernels from 3 to 1 → Significant speed improvement but we may lose accuracy
2. Tested intermediate configuration with all kernels but reduced C and gamma → Good balance (yet 1+ min for 500 samples)
3. Validated performance with different sample sizes to ensure scalability

---

Desired Solution

Optimize the hyperparameter ranges to achieve processing times under 5 seconds while maintaining acceptable accuracy for gaze prediction.