Learning Object Detectors from Weakly-Labeled Internet Images

In Proceedings 35th OAGM/AAPR Workshop, Graz, Austria

Learning visual object detectors typically requires a large amount of labeled data, which is hard to obtain. To overcome this limitation, we propose a three-stage system that avoids any human labeling and autonomously learns an object detector from unlabeled Internet images. In the first stage, we collect images via visual image search, just using the name of an object class. Then, in the second stage, we determine the presence of the target object and, finally, in the third stage, we estimate its localization and crop patches, which are used to learn a detector. Since we have to cope with ambiguously/wrongly labeled data, we apply multiple instance learning (MIL) techniques in the last two stages. In the experimental results, we demonstrate the benefits of the approach on publicly available benchmark datasets. In fact, we show that we can train competitive object detectors without using visually labeled data.

PDF