Uncertainty Guided Pedestrian Metadata Refinement Using Geometrical Representation
Author
Amin, Md Ibtihaj
Term
4. semester
Publication year
2026
Submitted on
2026-10-04
Abstract
This master’s thesis investigates how uncertainty can be used to improve the reliability of pedestrian metadata in image and video analysis. Pedestrian Attribute Recognition (PAR) models identify multiple visual attributes for a person, such as clothing, accessories, age-related properties and body orientation, but high overall performance does not guarantee that each individual attribute prediction is trustworthy. Under challenging conditions with occlusion, low resolution or poor lighting, the model may produce incorrect predictions with high confidence scores, exposing a gap between accuracy and reliability. The thesis analyzes two uncertainty estimation methods – predictive entropy and Monte Carlo dropout – at the attribute level to assess how well they distinguish between confident and potentially erroneous or ambiguous predictions. This per-attribute uncertainty is then used in a gating mechanism that selects only the most uncertain attributes for further verification by a Vision Language Model (VLM), which acts as a secondary source of visual semantic evidence. The VLM is thus used selectively, as an additional verification stage to revise and refine the original pedestrian metadata in a two-step pipeline, rather than predicting all attributes from scratch. In addition, the thesis examines how the geometry of the feature representation – conventional Euclidean space versus hyperbolic space – influences relationships between attributes, model behavior and the resulting uncertainty estimates. Euclidean and hyperbolic representations are evaluated with different visual backbones (including ResNet-50 and TinyViT) to explore the link between representation geometry, attribute-level uncertainty and metadata reliability. Although specific numerical results are not available in the provided excerpt, the work proposes an integrated framework that combines PAR, uncertainty analysis and VLM-based refinement to enhance the quality and trustworthiness of pedestrian metadata in applications such as surveillance, person retrieval and visual analytics.
Denne specialeafhandling undersøger, hvordan usikkerhed kan bruges til at forbedre pålideligheden af metadata om fodgængere i billed- og videoanalyse. Pedestrian Attribute Recognition (PAR) modeller identificerer flere visuelle attributter for en person, såsom beklædning, accessoires, alder og kropsorientering, men høje præcisionsmål for den samlede model betyder ikke nødvendigvis, at hver enkelt attribut er pålidelig. Særligt under udfordrende forhold med occlusion, lav opløsning eller dårlig belysning kan modellen give forkerte forudsigelser med høje scores. I afhandlingen analyseres to usikkerhedsmål – prædiktiv entropi og Monte Carlo dropout – på attributniveau for at vurdere, hvor godt de skelner mellem sikre og potentielt fejlbehæftede forudsigelser. Denne usikkerhedsinformation bruges i en udvælgelsesmekanisme, som kun sender de mest usikre attributter videre til en Vision Language Model (VLM), der fungerer som sekundær kilde til visuel semantisk evidens. VLM’en anvendes dermed selektivt som et ekstra verifikationslag til at revidere og raffinere de oprindelige metadata i en totrins-pipeline, i stedet for at forudsige alle attributter fra bunden. Afhandlingen undersøger desuden, hvordan den geometriske repræsentation af visuelle features – traditionel euklidisk geometri versus hyperbolsk geometri – påvirker relationer mellem attributter, modeladfærd og de resulterende usikkerhedsskøn. Euklidiske og hyperbolske repræsentationer evalueres med forskellige visuelle rygknogler (bl.a. ResNet-50 og TinyViT) for at belyse sammenhængen mellem repræsentationsgeometri, attributniveau-usikkerhed og metadata-pålidelighed. De konkrete kvantitative resultater fremgår ikke af det givne uddrag, men afhandlingen bidrager med et samlet rammeværk, der kombinerer PAR, usikkerhedsanalyse og VLM-baseret raffinering for at øge kvaliteten af fodgængermetadata i anvendelser som overvågning, persongenfinding og visuel analyse.
[This abstract has been generated with the help of AI directly from the project full text]
Keywords