Recently, object detectors based on deep learning have become widely used for vehicle detection and contributed to drastic improvement in performance measures. However, deep learning requires much training data, and detection performance notably degrades when the target area of vehicle detection (the target domain) is different from the training data (the source domain). To address this problem, we propose an unsupervised domain adaptation (DA) method that does not require labeled training data, and thus can maintain detection performance in the target domain at a low cost. We applied Correlation alignment (CORAL) DA and adversarial DA to our region-based vehicle detector and improved the detection accuracy by over 10% in the target domain. We further improved adversarial DA by utilizing the reconstruction loss to facilitate learning semantic features. Our proposed method achieved slightly better performance than the accuracy achieved with the labeled training data of the target domain. We demonstrated that our improved DA method could achieve almost the same level of accuracy at a lower cost than non-DA methods with a sufficient amount of labeled training data of the target domain.

Abstract:Recently, deep learning techniques have had a practical role in vehicle detection. While much effort has been spent on applying deep learning to vehicle detection, the effective use of training data has not been thoroughly studied, although it has great potential for improving training results, especially in cases where the training data are sparse. In this paper, we proposed using hard example mining (HEM) in the training process of a convolutional neural network (CNN) for vehicle detection in aerial images. We applied HEM to stochastic gradient descent (SGD) to choose the most informative training data by calculating the loss values in each batch and employing the examples with the largest losses. We picked 100 out of both 500 and 1000 examples for training in one iteration, and we tested different ratios of positive to negative examples in the training data to evaluate how the balance of positive and negative examples would affect the performance. In any case, our method always outperformed the plain SGD. The experimental results for images from New York showed improved performance over a CNN trained in plain SGD where the F1 score of our method was 0.02 higher.

No abstract

No abstract

The authors wish to make the following corrections to this paper [...]