CiPSI and CiPSINet: a New Dataset and Network for Finding Cars in Panchromatic Satellite Images

The low availability of high quality, labeled datasets of panchromatic satellite imagery featuring passenger vehicles for both academic and commercial applications led to the creation of CiPSI, a new dataset to fill exactly that gap. Using this new dataset, we evaluated six existing state-of-the-art classification networks and three detection methods. Ultimately, we introduce a novel network architecture, CiPSINet, that achieves higher performance than all other methods for vehicle classification while using fewer learnable parameters. Further, using this network, we improve the detection accuracy of the Faster R-CNN method by replacing the ResNet-50 backbone with CiPSINet.