Performance Evaluation of Bottom-Up Saliency Models for Object Proposal Generation
Anton Knaub, Vikram Narayan, Markus Adameck · 2016
In this work, we present the performance evaluation of twelve existing saliency models for the purpose of object proposal generation. The topic of object proposal generation is of high importance as most of the successful object detection techniques still employ the sliding window approach. Intelligent pre-selection of sliding windows using object proposal generation methods have shown to boost both runtime and detection performance. Unlike existing approaches, we do not employ any machine learning classifier or require training. The saliency map of the corresponding image is computed and the variance of the region of interest on the saliency map is considered as its score correspondingly. Experiments on two contemporary object detection datasets have revealed the effectiveness of bottom-up saliency models as a tool for object proposal generation.