Uncertainty-Aware Pedestrian Attribute Recognition via Evidential Deep Learning

Problem
This paper addresses the gap in pedestrian attribute recognition (PAR) systems regarding their inability to assess prediction reliability, particularly in the presence of low-quality samples. Existing deterministic methods lack mechanisms to quantify uncertainty, which can lead to unreliable predictions in complex real-world scenarios. The authors propose UAPAR, the first Evidential Deep Learning (EDL)-based framework for PAR, which aims to enhance robustness by effectively identifying unreliable predictions. This work is presented as a preprint and has not yet undergone peer review.

Method
UAPAR integrates EDL into a CLIP-based architecture, leveraging a Region-Aware Evidence Reasoning module that employs cross-attention mechanisms and spatial prior masks to extract fine-grained local features from pedestrian images. The architecture includes an evidence head that estimates attribute-wise epistemic uncertainty, allowing the model to quantify the reliability of its predictions. To further bolster training robustness, the authors introduce an uncertainty-guided dual-stage curriculum learning strategy, which mitigates the impact of severe label noise during training. The model’s architecture and training strategies are designed to enhance the interpretability and reliability of predictions in challenging scenarios.

Results
UAPAR demonstrates competitive or superior performance across multiple benchmarks, including PA100K, PETA, RAPv1, and RAPv2 datasets. The authors report significant improvements in accuracy metrics compared to baseline models, although specific numerical results are not detailed in the abstract. Qualitative evaluations indicate that UAPAR effectively generates uncertainty estimates that correlate with challenging or erroneous samples, suggesting that the model can discern between reliable and unreliable predictions.

Limitations
The authors acknowledge that while UAPAR improves robustness against label noise, the framework may still be sensitive to other forms of noise or adversarial attacks that were not explicitly tested. Additionally, the computational overhead introduced by the EDL components and the dual-stage curriculum learning may limit the model’s scalability in real-time applications. The paper does not address the potential impact of varying dataset characteristics on the model’s performance, nor does it explore the generalizability of the approach to other domains beyond pedestrian attribute recognition.

Why it matters
The introduction of uncertainty quantification in pedestrian attribute recognition has significant implications for downstream applications, such as autonomous driving and surveillance systems, where understanding the reliability of predictions is crucial for safety and decision-making. By providing a framework that not only predicts attributes but also assesses their reliability, UAPAR paves the way for more robust and interpretable AI systems in complex environments. This work could inspire further research into EDL applications across various domains, enhancing the reliability of machine learning models in real-world scenarios.

Authors: Zhuofan Lou, Shihang Zhang, Fangle Zhu, Shengjie Ye, Pingyu Wang
Source: arXiv:2604.26873
URL: https://arxiv.org/abs/2604.26873v1