Existing person re-identification (Re-ID) works mostly consider a short-term search problem assuming unchanged clothes and personal appearance. However, in real-world we often dress differently across locations, time, dates, seasons, weather, and events. As a result, the existing methods are unsuitable for long-term person Re-ID with clothes change involved. Whilst there are several recent long-term Re-ID attempts, a large realistic dataset with clothes change is lacking and indispensable for enabling extensive study as already experienced in short-term Re-ID setting. In this work, we contribute a large, realistic long-term person identification benchmark. It consists of 178K bounding boxes from 1.1K person identities, collected and constructed over 12 months. Unique characteristics of this dataset include: (1) Natural/native personal appearance (e.g., clothes and hair style) variations: The clothes-change and dressing styles all are highly diverse, with the reappearing gap in time ranging from minutes, hours, and days to weeks, months, seasons, and years. (2) Diverse walks of life: Persons across a wide range of ages and professions appear in different weather conditions (e.g., sunny, cloudy, windy, rainy, snowy, extremely cold) and events (e.g., working, leisure, daily activities). (3) Rich camera setups: The raw videos were recorded by 17 outdoor security cameras with various resolutions operating in a real-world surveillance system for a wide and dense block. (4) Largest scale: It covers the largest number of (17) cameras, (1, 121) identities, and (178, 407) bounding boxes, as compared to alternative datasets. Our dataset and benchmark codes are available on https://github.com/PengBoXiangShang/deepchange.
updated: Wed Aug 04 2021 13:40:29 GMT+0000 (UTC)
published: Mon May 31 2021 03:35:00 GMT+0000 (UTC)