Conference Papers Year : 2022

Gaussian Balanced Sampling for End-to-End Pedestrian Detector

Yang Yang
  • Function : Author
  • PersonId : 1405930
Jun Li
  • Function : Author
Biao Hou
  • Function : Author
Bo Ren
  • Function : Author
Xiaoming Jiang
  • Function : Author
Jinkai Cheng
  • Function : Author
Licheng Jiao
  • Function : Author

Abstract

Recently, NMS-free detector has become a research hotspot to eliminate negative influences, while NMS-based detector mis-suppress objects in crowd scene. However, NMS-free may face the problem of sample imbalance that affects convergence. In this paper, Gaussian distribution is adopted to fit the distribution of the targets so that samples can be chosen according to it. And we propose Gaussian Balance Sampling strategy to balance positive and negative samples actively. Besides, a simple loss function, PDLoss, is proposed to eliminate duplicated matches on the label assignment procedure and increase training speed. In addition, by a novel Non-target Response Suppression method, the designed network can focus more on hard samples and improve model performance. With these techniques, the model achieved a competitive performance on the CrowdHuman dataset.
Fichier principal
Vignette du fichier
537972_1_En_34_Chapter.pdf (3.19 Mo) Télécharger le fichier
Origin Files produced by the author(s)

Dates and versions

hal-04666416 , version 1 (01-08-2024)

Licence

Identifiers

Cite

Yang Yang, Jun Li, Biao Hou, Bo Ren, Xiaoming Jiang, et al.. Gaussian Balanced Sampling for End-to-End Pedestrian Detector. 5th International Conference on Intelligence Science (ICIS), Oct 2022, Xi'an, China. pp.318-325, ⟨10.1007/978-3-031-14903-0_34⟩. ⟨hal-04666416⟩
9 View
2 Download

Altmetric

Share

More