Overview of our framework. Grid Dropout (GD): Dropout is applied to specific pixels in a grid-like fashion. Center Dropout (CD): A large contiguous area is specified for pixel dropout. Center Filter (CF): After selecting the central region, basic mean filtering is applied. Random Shuffle (RS): Regions are randomly selected for shuffling pixel combinations in blocks.
Typical sample performance in our method ablation study. Original is the direct output of the backbone network; +SSL adds self supervised pre training; +RS&GD adds specific data augmentation based on +SSL.
Supported by the Open Fund of Science and Technology on Parallel and Distributed Processing Laboratory ( PDL )