Evaluating Novel Mask-RCNN Architectures for Ear Mask Segmentation

TL;DR

This paper compares new Mask-RCNN architectures for ear segmentation, demonstrating improved performance over the existing model across multiple datasets, but no single model is best universally.

Contribution

It introduces and evaluates three novel Mask-RCNN variants for ear segmentation, advancing the state-of-the-art in ear biometric detection.

Findings

New models outperform the existing Mask-RCNN in AP scores

No single model is best across all datasets

The study highlights the variability in model performance

Abstract

The human ear is generally universal, collectible, distinct, and permanent. Ear-based biometric recognition is a niche and recent approach that is being explored. For any ear-based biometric algorithm to perform well, ear detection and segmentation need to be accurately performed. While significant work has been done in existing literature for bounding boxes, a lack of approaches output a segmentation mask for ears. This paper trains and compares three newer models to the state-of-the-art MaskRCNN (ResNet 101 +FPN) model across four different datasets. The Average Precision (AP) scores reported show that the newer models outperform the state-of-the-art but no one model performs the best over multiple datasets.