Unleashing the Power of Data Augmentation and Training Strategy for Light Field Salient Object Detection

Light field salient object detection method highlights the prominent objects by integrating an all-focus image and focal slices. Existing methods heavily engage in designing sophisticated architectures to excavate appearance and latent depth clues for multi-modal fusion. To the best of our knowledge, we are the first to unleash the power of data augmentation and training strategy in the task. Specifically, an inter-modal bidirectional copy-paste augmentation is suggested to copy and paste local grids between all-focus images and focal slices, avoiding the excessive corruption and promoting multi-modal sample diversity. In addition, a dynamic adjustment strategy of training samples is introduced to select the suitable training samples in different training stages based on the difficulty of the samples. In the earlier stage, more simple samples and less difficult ones are utilized, while in the later stage, vice versa. Experiments demonstrate the impressive performance only based on data augmentation and training strategy, instead of complex model.