CloudWalk re-identification technology extends facial biometric tracking with improved accuracy
CloudWalk has developed re-identification technology that can track a person by clothing, hairstyle, and posture, to extend and supplement facial recognition for public surveillance.
In testing against three major datasets, Market-1501, DukeMTMC-reID, and CUHK03, CloudWalk says its Cross-mirror tracking ReID technology outperformed competitor offerings from Alibaba, Tencent, Microsoft, and the Chinese Academy of Sciences. The cloud-based system far surpassed the previous industry best mAP (mean Average Precision) accuracy rate against the Market-1501 dataset, reaching 91.14 percent, according to the Google translation of a company announcement. CloudWalk’s system outscored six challengers in both mAP and Rank-1 accuracy against all three datasets.
The company also said the core algorithm’s speed has increased by ten times over the past year, and that it has created algorithm modules for pedestrian detection, tracking, and structuring which are widely used in commercial, security, transportation, and other deployments. In a retail scenario, the technology can provide merchants with information on user behavior to make more accurate business decisions, according to the announcement.
The improvements in re-identification performance are at least partly due to a new “Peer-based Pedestrian Multi-Grain Feature Extraction Theory,” proposed as an extension of the “Pedestrian Multi-granularity Feature Extraction Theory” previously developed by CloudWalk researchers.
CloudWalk raised $140 million last year to support its research and expansion, and the company is expected to lead the global market for facial recognition devices through 2025, when it will be worth more than $7 billion dollars a year. The company has also played a role in surveillance systems used in Xinjiang, where an ambitious biometric collection program is associated with the alleged detention of up to a million people in internment camps.
NEC recently announced the development of its own re-identification technology.