Detection on Difficult Small Objects Using Layer-wise Training Strategy

Training strategy is a common condition which has considerable impact on the performance of deep learning networks. In this work, we propose a layer-wise training strategy which contains layers freezing, hard examples training and loss changing during training. Training networks using layer-wise training strategy can transfer pre-trained weights to similar tasks more efficiently. We detail about the way of freezing different layers in different epochs, which is associated with better stability. We also use small objects as training dataset, and it improves detection capabilities of networks on difficult small objects. The study offers some important insights into neural networks training. We further analyze the advantage of layer-wise training strategy and prove it efficient in experiments. Our method achieves superior accuracy over Fast R-CNN and Mask R-CNN on MS COCO dataset. We also detect photos contains difficult small objects using different models, and models trained by our strategy prove more efficient.