Focal Stack Alignment Enhancement Network For Light Field Salient Object Detection
Ziyi Huang, Binbin Yan, Shuo Chen, Dongliang Wang, Lu Yang · 2024
Light field data has benefited salient object detection (SOD), but it can be challenging to handle samples with various numbers of focal slices. One of the main problems is that the required number of slices often differs from the number available in existing datasets. Replicating the available slices to address this discrepancy may disrupt slice alignment and increase information redundancy. To address this, we present a novel network for focal stack alignment enhancement called FAENet, which strategically selects only a few but representative slices as input to reduce the likelihood of slice misalignment. FAENet has two pivotal modules: Collaborative Feature Refinement (CFR) to enhance multi-modal features of light field data and Saliency Edge Bridge(SEB) module to fuse these features and integrate edge supervision. These modules fully exploit light field information as input slices decrease. Experimental results on four benchmark datasets demonstrate our network’s superiority over 10 state-of-the-art models.