Domain Adaptation Plan: LIPE V2 (MPIIGaze → Gaze360)

1. Executive Summary

The goal of this plan is to bridge the "Domain Gap" between controlled laboratory data (MPIIGaze) and in-the-wild data (Gaze360). Our baseline zero-shot evaluation showed an MAE of 24.3° on Gaze360. We aim to reduce this to **< 7.0°** while maintaining a CPU-optimized inference speed of > 400 FPS.

2. Technical Strategies & Predicted Impact

Phase	Strategy	Technical Mechanism	Predicted MAE
0	Baseline	Zero-shot from MPIIGaze	24.3°
1	Domain Randomization	Resolution Dropout + Noise Augmentation	18.5°
2	Structural Adaptation	AdaLN + Feature-Level Alignment (DANN)	12.5°
3	Expert Guidance	Multi-Domain Knowledge Distillation	8.5°
4	Geometry Constraints	Self-Supervised Consistency Loss	< 7.0°

3. Detailed Implementation Roadmap

Phase 1: Data Hardening (Augmentation & Preprocessing)

Resolution Dropout: Randomly downsample 16x16 patches to 8x8 during training to simulate camera distance. Use cv2.INTER_CUBIC to prevent upsampling artifacts.
Edge-Preserving Denoising: Integrate Bilateral Filter (cv2.bilateralFilter) to smooth homogeneous regions and reduce grain without blurring the critical iris boundaries.
Tuned CLAHE Preprocessing: Apply CLAHE with a conservative clipLimit (1.0 - 1.2) to enhance contrast in dark regions without amplifying noise into artifacts.
Interpolation Strategy: Force INTER_CUBIC for all resizing to preserve the "Gaze Gradient" in low-resolution conditions.
Geometric Jitter: Add Gaussian noise to MediaPipe landmarks (±2px) to simulate tracking instability.
Extreme Pose Augmentation: Use existing rotation logic to synthetically create head poses beyond ±20°.

Phase 2: Feature Alignment (DANN & AdaLN)

DANN (Domain-Adversarial Neural Network):
- Implement a Gradient Reversal Layer (GRL).
- Add a Domain Classifier branch to the fusion_mlp to detect the source dataset.
- Train the backbone to produce domain-invariant features.
AdaLN (Adaptive Layer Normalization):
- Switch standard LayerNorm to domain-aware normalization.
- Store separate affine parameters ($\gamma, \beta$) for MPII and Gaze360.

Phase 3: Cross-Domain Distillation

Multi-Domain Teacher: Use L2CS-Net (Pre-trained on Gaze360) to provide soft labels for Gaze360 images.
Hybrid Loss: Combine Adaptive Wing Loss (Hard labels) with KL-Divergence (Teacher soft labels) across both datasets.

Phase 4: Self-Supervised Consistency

Mirror Invariance: Enforce $Yaw(Image) = -Yaw(Flipped_Image)$.
Equivariance Loss: If an image is rotated by $\theta$, the predicted gaze must shift by $\theta$ in the 3D spherical space.

4. Architectural Modifications (`src/models/student.py`)

Add GRL (Gradient Reversal Layer) class.
Implement DomainClassifier head (2-layer MLP).
Update forward() to handle domain_label and return domain_logits.
Refactor geo_mlp to include AdaLN logic.

5. Success Criteria

Accuracy: MAE on Gaze360 Frontal subset < 7.0°.
Generalization: Maintain MAE on MPIIGaze < 5.0°.
Efficiency: Latency < 1.5ms on single-core CPU.

Created: 2026-06-04 Status: INITIALIZED