Viewpoint invariant subject retrieval via soft clothing biometrics

Abstract

As

On categorising gender in surveillance imagery

“…Relative attributes were introduced in [19] and first applied to soft biometrics in [20], using relative ordered annotations of psychologically grounded global and body traits. Comparative soft biometrics have since been employed to describe clothing [9], face [1] and for the first time, gender [14], finding ambiguity in even clear images. Therefore, gender uncertainty and ambiguity concepts are supported in both early [6] and recent [14] works, and are evident visually in Figures 1, 3 and 5.…”

“…Comparative annotation has been successfully applied to soft biometrics [20,9] and most recently gender [14,15,1]. However, previous approaches cannot distinguish between attribute ambiguity (Figure 1a) and image obscurity ( Figure 1b), while pairwise comparisons for relative labels are difficult to scale.…”

“…We are primarily motivated by comparative soft biometrics [20,14,1,9], but our work is also analogous to [24,7], which investigate similarity comparisons via crowdsourcing, with several important distinctions. Firstly, both [24,7] collect overall similarity annotations from a wide range of image subject matter, finding broad, basic-level categories.…”

Analysing comparative soft biometrics from crowdsourced annotations

“…We further this work, by analysing soft trait label distributions, the affect of varying numbers of comparisons on inferred measurements and their correspondence to categorical labels. Jaha et al also augments these body annotations with soft clothing biometrics, enhancing recognition performance [14].…”

“…Combining several soft biometrics modalities, especially clothing, has proven important in improving subject recognition rates [3,14] and can be estimated for surveillance tracking and search [9,44].…”