Diverse receptive field network with context aggregation for fast object detection

摘要

Current context-utilizing detectors are all based on two-stage approaches. However, their computational efficiency and context quality extremely depend on the accuracy of proposal-generating methods, which limits their performance and makes them hardly perform real-time detection. In this work, we present a context-exploited method that integrates features in different receptive fields to obtain contextual representation. Based on this idea, we put forward the multi-branch diverse receptive field modules (DRF modules) and their design principles to encode context. To further utilize contextual information for fast object detection, we propose a one-stage diverse receptive field network (DRFNet). In DRFNet, the DRF modules are first applied to capture rich context as the basis, then a parallel structure is constructed to exploit the context at different scales along with DRF modules. Comprehensive experiments indicate that the context introduced by our methods improves the detection performance and DRFNet achieves a good trade-off between speed and accuracy. (C) 2020 Published by Elsevier Inc.