Engineering
ML Tool for Hand Landmark Detection
Developed an automated machine learning pipeline to replace the UF Musculoskeletal Biomechanics Lab’s slow, manual ImageJ processing of hand images, enabling rapid, consistent extraction of 31 standardized hand-landmark measurements from an original dataset of 495 hand images.
My Role
Independent Researcher & ML Developer
Duration
1 yr 8 months
Tools
Python, TensorFlow/Keras, MobileNetV2 architecture
Link
Overview
The 1987–1988 Anthropometric Survey of U.S. Army Personnel used outdated, limited‐demographic hand measurements, and the lab’s reliance on manual ImageJ workflows was time-consuming and prone to human error, hindering large-scale, inclusive biomechanical studies.
I adapted a facial landmark detection model (MobileNetV2 architecture) for hand landmark detection by building a dataset of manually annotated hand images, preprocessing each with orientation correction and color thresholding, and fine-tuning the network’s width multiplier (α) to maximize landmark localization accuracy.
Research
Conducted an in-depth literature review on MobileNetV2’s architecture and transfer learning best practices, consulting published papers and ML experts to identify which hyperparameters (like the width multiplier α) to experiment with.
Mapped the original U.S. Army Personnel survey landmark numbering to our annotation scheme by creating a detailed key that aligned each measurement description to its corresponding image landmark, ensuring accurate distance calculations in downstream computation and error analysis.
Design
Organized the 495 annotated images into four orientation folders based on ruler position (bottom/top/left/right), applied an OpenCV‐based pipeline for color-threshold detection and rotation rules, excluded top-oriented images (final dataset included 408 images), removed ruler landmarks, and split data 70/15/15. All preprocessing and initial training occurred in Jupyter Notebook; subsequent runs used Colab’s T4 GPU. Models trained for 30 epochs (batch size 4) with MSE loss and Adam optimizer, saving the best checkpoint via early stopping.
Results
Reduced per-image landmark extraction time from minutes to milliseconds.
Demonstrated 79.55% average detection accuracy, validating AI-driven anthropometry.
Established a scalable framework for expanding dataset diversity and further model refinement.
