Automated Training Set Size Reduction for Detection of Small and Thin Objects

Convolutional neural network-based object detectors such as traffic sign detectors can achieve high accuracy, but are computationally expensive to train. This becomes an acute problem as the size of training sets increase. Thus we need methods to eliminate (1) redundant data in the training sets to reduce the training time and (2) mislabelled data to increase accuracy. Recent works have shown that it is possible to decrease training time, but struggle in improving accuracy. Therefore, we propose three methods for training set size reduction with the aim to improve the detector’s accuracy measured in intersection over union. The first method uses label information to remove images from the training set: we discard images without objects of interest. In contrast, the second method uses feature data of the input images: we discard images that are too similar. Finally, the third method combines the two above-mentioned methods.