|
Publication:Hong Mo, Kaige Li, Wei Wu, Localization in Crowds with Loosen-Structured Pair-of-points Loss[C]. IEEE International Conference on Virtual Reality and Visualization (ICVRV)，Kunming, China, 18-20 November 2022. pdf (CCF-C conference) |
| |
Phantom dots are common spatial semantic errors in crowd localization and counting. We attribute these semantic errors to the absence of bounding boxes and deal with them through adjacent nodes. One approach to solving this problem involves constraining the point and its neighbor point set simultaneously. However, learning two point sets that satisfy the subjective relation is challenging because predicting a point set is an open-set problem. In this paper, we propose a loosen structured pair-of-points loss, making predicting a point set and its neighbor point set simultaneously possible. On the other hand, to suppress the probability of points far away from the target points, we compose the connected component map, which is free
with head size. Without bells and whistles, experiment results on several public datasets reach the state-of-the-art.
|
| create by admin at 2023-01-19 21:40:30 |
|