Head Pose Experiments

Snapshot

• Period: Local experiment folder present in recent workspace; no Git history detected
• Source: `/Users/jose/Developer/work/sparshiq/experiments/head-pose`
• Domain: Head pose estimation, computer vision research survey
• Status: Exploration/reference integration

Portfolio Summary

This folder appears to be a research/evaluation workspace rather than a committed project. It collects several head-pose estimation approaches and supporting dataset tooling: DirectMHP for direct multi-person full-range head pose, 6DRepNet360 for robust 360-degree rotation estimation, CMU Panoptic utilities for 3D keypoint reprojection, and lightweight ResNet/MobileNet head-pose models with ONNX export/inference paths.

Stack

• Python
• PyTorch
• OpenCV-style computer vision utilities
• YOLO-style detection/training code
• ONNX export and inference paths
• CMU Panoptic dataset tooling
• 3D visualization/reprojection scripts

Potential Portfolio Angle

• Define a target use case, such as driver attention, classroom analytics, rail-cab operator monitoring, or multi-person scene analysis.
• Compare model families on the same clips or images.
• Export one lightweight model to ONNX and benchmark CPU/GPU inference.
• Build a small demo that overlays yaw/pitch/roll or 3D axes on video.
• Document tradeoffs: single-person vs multi-person, full-range vs frontal, accuracy vs model size, 2D detection coupling vs 3D representation.

Blockers To Document

• Dataset licensing and access constraints for AGORA, CMU Panoptic, 300W-LP, AFLW2000, and BIWI.
• Model zoo fragmentation across research repos.
• Different angle conventions and evaluation metrics across projects.
• Need to distinguish copied reference code from original integration/evaluation work.