Gaze Evaluation Platform

Evaluation infrastructure for driver-monitoring gaze models: a system that answers “which model should we ship, and how do we know?” automatically and repeatedly.

Instead of running a notebook once and eyeballing numbers, every model evaluation is containerized, logged, and compared automatically. Adding a new model requires writing one YAML config file and changing nothing else.

Core Workflow

1. Define a model in a YAML config (architecture, weights, quantization, runtime)
2. Run the evaluation runner — it loads the model, runs inference on preprocessed face crops, computes metrics, and logs everything to MLflow
3. Open a PR — CI builds the Docker image, runs the eval, and blocks the merge if any metric regresses
4. Compare models in the dashboard — a FastAPI service with an interactive Pareto frontier view

Metrics

Category	Metrics
Accuracy	MAE (yaw, pitch, combined), p50/p95/p99 angular error
Safety	Eyes-off-road false positive rate, false negative rate
Latency	Per-batch mean, p50, p95, p99 (ms)
Slicing	All accuracy metrics broken out by head-pose bin (0–15°, 15–30°, 30–45°, 45°+)

Hardware info (device, chip) is logged alongside latency so comparisons are apples-to-apples.

CI Regression Gate

Every pull request to main triggers a GitHub Actions workflow that builds the Docker image, runs evaluation on a test fixture, compares results against a baseline, and fails the build if any metric regresses beyond its tolerance — for example, MAE can’t increase by more than 0.5 degrees, FNR can’t increase by more than 2%.

Dashboard

A FastAPI service reads from MLflow and serves an interactive dashboard with a models overview, sortable run history, per-model detail pages, and a Pareto plot comparing any two metrics with Pareto-optimal models highlighted.

Tech Stack

Language: Python Preprocessing: MediaPipe BlazeFace detection + crop pipeline Tracking: MLflow Data Versioning: DVC Dashboard: FastAPI, Tailwind CSS, Chart.js Infrastructure: Docker, GitHub Actions

Links

• GitHub