Existing crowd counting methods have generally used Gaussian smoothing for correct data represented by dots, but they have shown experimentally/theoretically that this has a negative impact on the generalization performance. They also proposed a loss function based on the optimal transport and SOTA on various datasets.
TL;DR
Existing crowd counting methods have generally used Gaussian smoothing for correct data represented by dots, but they have shown experimentally/theoretically that this has a negative impact on the generalization performance. They also proposed a loss function based on the optimal transport and SOTA on various datasets.
Why it matters:
Paper URL
https://arxiv.org/abs/2009.13077
Submission Dates(yyyy/mm/dd)
2020/09/28
Authors and institutions
Boyu Wang, Huidong Liu, Dimitris Samaras, Minh Hoai
Methods
Results
Comments