BROWSE
Detail
Conference Paper
보행자 시점의 교차로 인식을 위한 이미지넷 선학습 모델의 중요성
Cited - time in 
Share 
Share
- Authors
- Marcella Astrid, Muhammad Zaigham Zaheer, 이승익
- Issue Date
- 2021-07
- Citation
- 대한전자공학회 학술 대회 (하계) 2021, pp.1565-1568
- Publisher
- 대한전자공학회
- Language
- Korean
- Type
- Conference Paper
- Abstract
- Pedestrian-view intersection classification is a key component which assists the navigation of autonomous small robots. Previous works finetune ImageNet-pretrained network with the dataset. In this work, we conduct experiments to analyze the importance of fine-tuning from ImageNet-pretrained network by comparing it with training from scratch. In order to compensate the additional training during the pretraining, we also train the model from scratch using higher learning rate and/or more training epochs. Despite that, model trained from ImageNet-pretrained network still shows significantly higher performance compared to the model trained from scratch. This demonstrates the importance of pretraining in pedestrian-view intersection classification.
- KSP Keywords
- Fine-tuning, Higher Learning, Higher performance, Key Components, Learning rate, Training Epochs