| # Model Complexity Analysis: LIPE V2 Gaze Estimation |
|
|
| This document provides a definitive technical breakdown of the architectural complexity for the LIPE V2 Student model across different configurations. These metrics have been verified using automated code analysis (`src/utils/verify_complexity.py`). |
|
|
| ## 1. Comparative Complexity Matrix |
|
|
| | Metric | 8x8 Baseline | 16x16 Baseline (Current) | **Gold Standard (16x16 + Fusion)** | |
| | :--- | :---: | :---: | :---: | |
| | **Input Resolution** | 8 x 8 px | 16 x 16 px | 16 x 16 px | |
| | **Fusion Type** | Residual Addition (`+`) | Residual Addition (`+`) | **Concatenation + MLP** | |
| | **Total Parameters** | **379,188** | **379,188** | **510,516** | |
| | **Total FLOPs** | **1.93 MFLOPs** | **20.99 MFLOPs** | **21.25 MFLOPs** | |
| | **Model Size (.pt)** | ~1.45 MB | ~1.45 MB | ~1.95 MB | |
| | **Inference (CPU)** | ~0.8 ms | ~1.1 ms | ~1.2 ms | |
| | **Max FPS** | ~1250 | ~900 | ~830 | |
|
|
| ## 2. Parameter Derivation (Layer-by-Layer) |
|
|
| | Module | Layer | Configuration | Calculation | Parameters | |
| | :--- | :--- | :--- | :--- | :--- | |
| | **Appearance (CNN)** | Conv1 | 1 $\to$ 16, 3x3 | (1 * 3 * 3 * 16) + 16 | 160 | |
| | | Conv2 | 16 $\to$ 32, 3x3 | (16 * 3 * 3 * 32) + 32 | 4,640 | |
| | | Conv3 | 32 $\to$ 64, 3x3 | (32 * 3 * 3 * 64) + 64 | 18,496 | |
| | **Geometric (MLP)** | Geo_MLP1 | 956 $\to$ 256 | (956 * 256) + 256 | 244,992 | |
| | | LayerNorm | 256 | (256 * 2) | 512 | |
| | | Geo_MLP2 | 256 $\to$ 256 | (256 * 256) + 256 | 65,792 | |
| | **Fusion (Gold Only)** | Fusion_MLP | 512 $\to$ 256 | (512 * 256) + 256 | 131,328 | |
| | **Heads** | Pitch_Head | (256, 64, 90) | (256*64+64) + (64*90+90) | 22,298 | |
| | | Yaw_Head | (256, 64, 90) | (256*64+64) + (64*90+90) | 22,298 | |
| | **TOTAL** | | | | **510,516** | |
| |
| ## 3. Key Architectural Observations |
| |
| ### Adaptive Stability |
| The use of **Adaptive Average Pooling** ensures that the convolutional feature maps are flattened into a fixed 64-dimensional vector regardless of the input resolution (8x8 or 16x16). This design choice keeps the total parameter count constant when scaling resolution, preventing model "bloat." |
| |
| ### Real-Time Efficiency |
| Even in the highest-fidelity configuration (**Gold Standard**), the model utilizes only **~17.7%** of the targeted computation budget (120 MFLOPs). This leaves significant thermal and computational headroom for other background tasks on consumer-grade hardware. |
| |
| ### Fusion Scalability |
| Upgrading to **Concatenation Fusion** increases the parameter count by **~34%** but only increases the computational load (FLOPs) by **< 1%**. This makes it an extremely efficient method for improving model accuracy without sacrificing real-time performance. |
| |
| --- |
| *Generated and Verified on: 2026-06-02* |
| |
