School of Engineering Department of Computer Science and Engineering 91 AI Meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: CHEN I Chieh / COMP Course: UROP1100, Spring Crowd counting has been an important technique in urban planning and public safety, and it has been a popular field in artificial intelligence. Initial works had tried to solve the problem with detection-based object counting approaches; however, these methods would suffer from data with dense crowd or objection occlusion. Recent works have been proposing regression-based methods and have achieved success and overcome the hazard above. In this project, we are going extend this problem to handle nighttime crowd data with existing crowd counting models and discuss possible way to improve the models’ performance on nighttime data. We will focus this problem particular at the domain of data processing and further learning or modification on existing models. AI Meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: DAI Tianyuan / COSC Course: UROP2100, Fall Crowd counting is to estimate the number of target objects (e.g., person, cars) in crowd-scene images or frames of videos. Current state-of-art methods depend heavily on supervised learning but the data is scarce. It adopts the assumption that all data should be identically and independently distributed (a.k.a., i.i.d.). However, data in the world often violate this i.i.d. assumtion, having different scales and distributions. This induces super difficulty in transferring crowd counting models trained on a particular dataset to other datasets. Therefore, how to do Out-of-Distribution (OoD) generalization in crowd counting to boost the performance of crowd counting models across different datasets becomes a great demand. In this paper, two main methods are investigated to do Out-of-Distribution generalization in crowd counting tasks: 1) through semi-supervised/unsupervised learning; 2) through dataset construction. An experiment is conducted to show that our dataset outperform current datasets, which shows that it could be adopted to boost the performance of crowd counting models across different scenes and datasets.
RkJQdWJsaXNoZXIy NDk5Njg=