You Only Cut Once: Boosting Data Augmentation with a Single Cut

In order to boost the generalization functionality and robustness of neural networks in personal computer vision, information augmentation tactics have been proposed. A current review posted on indicates a tactic of doing augmentations beyond the impression amount.

Graphic credit score: Unfavorable House, CC0 General public Domain

Researchers propose to cut a person picture into two equal items, execute facts augmentations separately inside of each individual piece, and concatenate the two augmented parts again jointly. This sort of an tactic increases the range at the regional level as properly as at the holistic image degree. It also encourages neural networks to share the exact same cognitive skill to recognize objects from partial information and facts as people can.

Substantial evaluations confirm that the proposed technique exhibits constructive final results in a variety of laptop or computer eyesight duties and datasets. Scientists make a conclusion that undertaking details augmentations can be a core element in schooling neural networks.

We present You Only Slice After (YOCO) for executing information augmentations. YOCO cuts one image into two pieces and performs details augmentations independently inside of each piece. Applying YOCO increases the diversity of the augmentation for every sample and encourages neural networks to recognize objects from partial facts. YOCO enjoys the qualities of parameter-no cost, easy utilization, and boosting pretty much all augmentations for no cost. Comprehensive experiments are done to examine its effectiveness. We very first reveal that YOCO can be seamlessly used to varying details augmentations, neural community architectures, and provides effectiveness gains on CIFAR and ImageNet classification tasks, sometimes surpassing common graphic-amount augmentation by big margins. What’s more, we exhibit YOCO gains contrastive pre-education towards a additional highly effective illustration that can be superior transferred to various downstream jobs. At last, we review a variety of variants of YOCO and empirically review the performance for respective settings. Code is readily available at GitHub.

Investigation paper: Han, J., “You Only Lower As soon as: Boosting Knowledge Augmentation with a One Cut”, 2022. Hyperlink: