FootNet — Multi-View Human Feet Image Dataset for Foot Segmentation

FootNet is a multi-view dataset of human feet images for foot segmentation, acquired via consumer smartphones in realistic outpatient clinical conditions. It contains 191 image–mask pairs organised into six standardised anatomical views (left/right × side/top/sole), with binary foreground masks manually drawn in LabelMe by a qualified clinician. Images were captured hand-held under unconstrained ambient lighting and against varied backgrounds — the setting a deployed point-of-care tool would actually encounter.

The release includes the dataset, a trained DeepLabV3-MobileNetV3-Large checkpoint (best validation Dice 0.9613, ~127 MB), per-image SAM (ViT-B) zero-shot evaluation results, the training script, and the inference pipeline. EXIF metadata has been stripped from all images and visible payment-card details manually redacted.

Position: a seed dataset to support reproducible research on foot segmentation in unconstrained smartphone imagery.