Benchmark Analysis of Backbone Network Architectures for Domain Adaptive Object Detection

Complexity of the convolutional neural networks, which is the main component of different applications such as classification, object detection and segmentation, is increased to obtain better performance. The increase in model complexity leads to dramatic increase in inference time and memory usage (number of parameters). Mobile networks which are developed to overcome these issues, achieved successful results. Therefore, in terms of performance, the recent mobile networks are considered to be comparable to complex network architectures despite their simple structure. In this paper, the performance of mobile networks in the field of domain adaptive object detection applications are examined. Domain Adaptive Faster R-CNN is chosen as the baseline method and various mobile networks are integrated into the backbone of the model. All experiments are carried out on the Cityscapes dataset. The results are compared in terms of mean average precision, number of parameters, and number of processes.