Soft Biometric Recognition From Comparative Crowdsourced Annotations

Abstract

Soft biometrics provide cues that enable human identification from low quality video surveillance footage. This paper discusses a new crowdsourced dataset, collecting comparative soft biometric annotations from a rich set of human annotators. We now include gender as a comparative trait, and find comparative labels are more objective and obtain more accurate measurements than previous categorical labels. Using our pragmatic dataset, we perform semantic recognition by inferring relative biometric signatures. This demonstrates a practical scenario, reproducing responses from a video surveillance operator searching for an individual. The experiment is guaranteed to return the correct match in the the top 7% of results with 10 comparisons, or top 13% of results using just 5 sets of subject comparisons.

Soft biometric retrieval to describe and identify surveillance images

“…Earlier soft biometric studies focus on annotating a set of absolute categorical labels, analysing the mutual exclusivity and identifying ability of each soft trait [25,5]. Later studies collect comparative annotations, by comparing pairs of subjects in a more objective manner [24,16,11]. Our dataset and investigation incorporates both forms of annotation.…”

“…In order to classify image attributes relative to one another, Parikh et al proposes an extension to the RankSVM algorithm including similarity constraints [20]. As a consequence, many studies use this algorithm to perform subject ranking given pairwise comparisons [17,11], and it is used to model our dataset's relative continuous and relative binary labels, following Martinho et al [16].…”

“…The second set of annotations are provided by Martinho et al, collecting pairwise subject comparisons, through an online crowdsourcing task using a 5-point, bi-polar scale [16]. Relative continuous values are derived from the pairwise annotations using the RankSVM formulation with similarity constraints, detailed by Parikh et al [20].…”

“…Soft traits (e.g gender, height etc.) are derived from salient and permanent, global and body characteristics [19,25,24,16], setting them apart from other mid-level semantic attributes.…”

“…We present a novel dataset with which to investigate several re-identification scenarios, including both images [26] ( Figure 1) and a comprehensive set of soft biometric annotations [25,16]. Other widely used benchmark datasets (e.g.…”