Feature Selection for Subject Ranking using Soft Biometric Queries

Abstract

This paper presents a feature selection model that aims to identify subjects from low-resolution surveillance images based on a soft biometric description query. The process is divided into three main stages. In the first stage, semantic segmentation is performed on the subjects, classifying and localising different parts of their bodies / accessories. The second stage extracts information from the segmentations and maps each subject to a vector in a soft biometric feature space. Last but not least, the purpose of the final stage is to find a good weighting on the features extracted in the previous step, based on the intuition that some of them are more important, more accurate or have a higher variance. It is assumed that the matching process might benefit considerably from a set of good weights. Analysis on the IEEE AVSS Challenge dataset shows encouraging performance for segmentation and subject matching with the correct subject reliably matched just outside the top ten on the training set, and just outside top 10% on the recently released test set.

A Quadruplet Loss for Enforcing Semantically Coherent Embeddings in Multi-Output Classification Problems

“…Several works relied regarded semantic segmentation as a tool to support labels inference: Galiyawala et al [10] described a deep learning framework for person retrieval using the height, clothes' color, and gender labels, with a semantic segmentation module used to remove clutter. Similarly, Cipcigan and Nixon [3] obtained semantically segmented regions of the body, that subsequently fed two CNN-based feature extraction and inference modules.…”

“…Here, our goals were essentially to compare the performance of the quadruplet loss with respect to the baselines and also to perceive any variations in performance with respect to different learning architectures. A TensorFlow implementation of the of both architectures is available at 3 .…”