PDL3D: 3D Attention Module with Partial Dense Layer for Small-to-Medium Dataset on Object Detection

Deep learning typically requires a large amount of training data. In this paper, we propose a PDL3D module, which exploits both channel attention and spatial attention mechanisms to improve the performance of deep convolutional neural networks (CNNs). PDL3D is a generic module that can be inserted into any CNN architecture and can be trained end-to-end with the inserted CNN architecture. Following the concept of MobileNet, PDL3D incurs less computation complexity in spatial attention. We prove it to be helpful in handing small to medium datasets by dividing MS COCO into smaller datasets, which we call mini coco datasets, and validating PDL3D on them with extensive experiments. Finally, we test it on a real PCB (Printed Circuit Board) dataset from electronic industry. Our experiments show that training PDL3D with small-to-medium datasets achieves similar or better performance compared to training existing networks with large datasets. Several CNN backbones have been tested to validate our claims.