Improving Deep Object Detection Backbone with Feature Layers

摘要

Deep neural networks are the frontier in object detection, a key modern computing task. The dominant methods involve two-stage deep networks that heavily rely on features extracted by the backbone in the first stage. In this study, we propose an improved model, ResNeXt101S, to improve feature quality for layers that might be too deep. It introduces splits in middle layers for feature extraction and a deep feature pyramid network (DFPN) for feature aggregation. This backbone is neither much larger than the leading model ResNeXt nor increasing computational complexity distinctly. It is applicable to a range of different image resolutions. The evaluation of customized benchmark datasets using various image resolutions shows that the improvement is effective and consistent. In addition, the study shows input resolution does impact detection performance. In short, our proposed backbone can achieve better accuracy under different resolutions comparing to state-of-the-art models.