Abstract

Recently many efficient object detectors based on convolutional neural networks (CNN) have been developed and they achieved impressive performance on many computer vision tasks. However, in order to achieve practical results, CNNs require really large annotated datasets for training. While many such databases are available, many of them can only be used for research purposes. Also some problems exist where such datasets are not available at all or they are limited in scope, e.g. many robotics applications. However, it is usually possible to obtain a large set of unlabelled data which contain useful information. The above mentioned possibility justifies a development of methods that exploit unlabelled data and they work with a minimal number of required annotations. In this work we follow recent self-supervised learning paradigm. Large unlabelled dataset of traffic monitoring was acquired by the authors. Then CNN was trained in order to perform moving objects segmentation based on labels obtained from a unsupervised motionbased segmentation algorithm. Even though collected labels are not perfect, they still allow CNN to learn an efficient feature representation. In the next step we fine-tuned the CNN algorithm on a limited set of manually labelled ground-truth data for object detection. Subsequently, we investigated the relation between the number of labels used for fine-tuning and final detection performance on the test set. We also compared the results with CNN pretrained on ImageNet which is now a common technique. Vehicle detection results obtained on our custom dataset are presented in the paper. The obtained results are promising, because they demonstrate that even when only limited ground-truth data are available, it is still possible to learn efficient feature representation given large collection of unlabelled data. The presented approach seems applicable to any object detector setting where there is an access to a large set of unlabelled data with moving objects of interest.

Authors (2)