Domain Adaption of Vehicle Detector based on Convolutional Neural Networks

Generally the performance of a vehicle detector will decrease rapidly, when it is trained on a fixed training set but applied to a specific scene with view changes. The reason is that in the training set only a few samples are helpful for vehicle detection in the specific scene while other samples disturb the accurate detections. To solve this problem, we propose a novel transfer learning method to adapt the trained vehicle detector based on convolutional neural networks (ConvNets) to a specific scene with several new labeled samples. At first we reserve the share-filters and update the non-shared filters to improve the sensitivity of the vehicles in the specific scene. Then we combine the similar feature maps to accelerate the detection speed. At last for making the vehicle detector stable, we fine-tune it several times with the updated training set. Our contributions are an original research on transferring the vehicle detector based on ConvNets and an optimization approach about removing the redundant connections in the ConvNet vehicle detector. The extensive comparative experiments on three different datasets demonstrate that the transferred detectors achieve the improvements on both of the accuracy and speed.