Rethinking Dynamic Scale Training with Regularization Effect of Data Augmentation and Data Pool

Natural data collected from the real world often exhibit the scale imbalance problem. A large object can produce much more loss values than a small object, causing the detector to favour large objects more, even though small objects dominant the dataset. This inclination inside detectors results in the performance degradation of small objects. To alleviate this problem, this paper proposes a new patch-level collage fashion data augmentation technique and a new global scheduler based on existing dynamic scale training paradigm. Our new data augmentation can generate collage images with uniform object scales for better augmentation effects. Additionally, our new global scheduler can adjust the strength between different data augmentations to adapt to different stages of the training process. Experiments demonstrate the effectiveness of our techniques. Codes at https://github.com/Andisyc/DataPool.