Semi-supervised Object Detection via Virtual Category Learning

Abstract

Due to the lack of large amounts of labelled data to learn rich-expressive features of objects, semi-supervised detectors powered by pseudo labelling techniques usually make a tentative decision for the pseudo labels of confusing samples. When dealing with confusing training samples, neither of the two recently adopted strategies, i.e., discarding or retaining, is optimal. Arbitrarily discarding the valuable confusing samples would compromise the generalisation ability of the model, while using them for model training would exacerbate the confirmation bias issue caused by inevitable mislabelling. To remedy this situation, this paper, for the first time, proposes to make use of these confusing samples without label correction, instead of discarding them. To this end, an alternative virtual category (VC), which is safe for optimisation, is provided for each confusing sample such that they can still contribute to better decision-boundary finding even without a concrete label. It is attributed to a new VC loss formulating the embedding distance between the training sample and the virtual category as the lower bound of the inter-class distance. Moreover, we also modify the localisation loss to allow the high-quality bounding boxes to be used for location regression training. Extensive experimental results demonstrate that the proposed semi-supervised detector underpinned by the VC learning surpasses the current state-of-the-art, when extremely small amounts of annotated labels are available for training (e.g., 0.5\% and 1\% label ratios).

Tasks

Reproductions