Shifting more attention to nature of saliency
Abstract: Saliency detection is usually defined as a point estimation problem, with a deterministic model trained to achieve regression from input space to output space. We argue that the deterministic mapping fails to explain the “subjective nature” or “uncertainty” of saliency. To this end, we propose the latent variable model-based saliency detection network with the latent variable to model the “subjective nature” of saliency, leading to stochastic predictions. Further, following conventional uncertainty estimation pipeline, we also explore ensemble-based methods for stochastic saliency prediction. With extensive experimental results on uncertainty from both the latent variable models and ensemble-based models, we observe more reasonable uncertainty estimation and more accurate prediction with the former techniques.
About the speaker: Jing Zhang is currently a Postdoc Fellow with School of Computing, the Australian National University. Her main research interests include saliency detection, weakly supervised learning, generative model. She won the Best Student Paper Prize at DICTA 2017, the Best Deep/Machine Learning Paper Prize at APSIPA ASC 2017 and the Best Paper Award Nominee at IEEE CVPR 2020.
Date and Time
Location
Hosts
Registration
-
Add Event to Calendar
- CANBERRA, Australian Capital Territory
- Australia 2601
- Contact Event Host
- Co-sponsored by Jue Zhang