Fraunhofer-Gesellschaft

Publica

Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Hardness Prediction for More Reliable Attribute-based Person Re-identification

 
: Florin, Lucas; Specker, Andreas; Schumann, Arne; Beyerer, Jürgen

:

Institute of Electrical and Electronics Engineers -IEEE-; IEEE Computer Society:
IEEE 4th International Conference on Multimedia Information Processing and Retrieval, MIPR 2021 : 8-10 September 2021, Virtual Event
Piscataway, NJ: IEEE, 2021
ISBN: 978-1-6654-4814-7
ISBN: 978-1-6654-1865-2
S.418-424
International Conference on Multimedia Information Processing and Retrieval (MIPR) <4, 2021, Online>
Englisch
Konferenzbeitrag
Fraunhofer IOSB ()
attribute recognition; re-id; cross modal; pedestrian; retrieval

Abstract
Recognition of person attributes in surveillance camera imagery is often used as an auxiliary cue in person re-identification approaches. Additionally, increasingly more attention is being payed to the cross modal task of person re-identification based purely on attribute queries. In both of these settings, the reliability of attribute predictions is crucial for success. However, the task attribute recognition is affected by several non-trivial challenges. These include common aspects, such as degraded image quality through low resolution, motion blur, lighting conditions and similar factors. Another important factor in the context of attribute recognition is, however, the lack of visibility due to occlusion through scene objects, other persons or self-occlusion or simply due to mis-cropped person detections. All these factors make attribute prediction challenging and the resulting detections everything but reliable. In order to improve their applicability to person re-identification, we propose to apply hardness prediction models and provide an additional hardness score with each attribute that measures the likelihood of the actual prediction to be reliable. We investigate several key aspects of hardness prediction in the context of attribute recognition and compare our resulting hardness predictor to several alternatives. Finally, we include the hardness prediction into an attribute-based re-identification task and show improvements in the resulting accuracy. Our code is available at https://github.com/Lucas-Florin/hardness-predictor-for-par.

: http://publica.fraunhofer.de/dokumente/N-642389.html