Closed gngdb closed 9 years ago
Worth linking the baidu paper here, recent paper with references on data augmentation (and actually mention how they performed the augmentation): http://arxiv.org/abs/1501.02876
Submitted a model with augmentation, closing this now.
The idea is pretty simple; we're trying to increase the number of training examples by generating new examples from modifications of the existing data. Intuitively, it seems like this task could work well with this, because the rotation and location of the plankton in each image is incidental.