Delving into the Effectiveness of Receptive Fields: Learning Scale-Transferrable Architectures for Practical Object Detection

Scale-sensitive object detection remains a challenging task, where most of the existing methods could not learn it explicitly and are not robust. Besides, they are less efficient during training or slow during inference, which is not friendly to real-time applications. In this paper, we propose a scale-transferrable architecture for practical object detection based on the analysis of the connection between dilation rate and effective receptive field. Our method firstly predicts a global continuous scale, which is shared by all positions, for each convolution filter of each network stage. Secondly, we average the spatial features and distill the scale from channels to effectively learn the scale. Thirdly, for fast-deployment, we propose a scale decomposition method that transfers the robust fractional scale into the combination of fixed integral scales for each convolution filter, which exploits the dilated convolution. Moreover, to overcome the shortcomings of our method for large-scale object detection, we modify the Feature Pyramid Network structure. Finally, we illustrate the orthogonality role of our method for sampling strategy. We demonstrate the effectiveness of our method on one-stage and two-stage algorithms under different configurations and compare them with different dilated convolution blocks. For practical applications, the training strategy of our method is simple and efficient, avoiding complex data sampling or optimization strategy. During inference, we reduce the latency of the proposed method by using the hardware accelerator TensorRT without extra operation. On the COCO test-dev, our model achieves 41.7% mAP on one-stage detector and 42.5% mAP on two-stage detector based on ResNet-101, and outperforms baselines by 3.2% and 3.1% mAP, respectively.