Label-Assemble: Leveraging Multiple Datasets with Partial Labels

The success of deep learning relies heavily on large and diverse datasets with extensive labels, but we often only have access to several small datasets associated with partial labels. In this paper, we start a new initiative, “LabelAssemble”, that aims to unleash the full potential of partially labeled data from an assembly of public datasets. Technically, we introduce a new dynamic adapter to encode different visual tasks, which addresses the challenges of incomparable, heterogeneous, or even conflicting labeling protocols. We also employ pseudo-labeling and consistency constraints to harness data with missing labels and to mitigate the domain gap across datasets. From rigorous evaluations on three natural imaging and six medical imaging tasks, we discover that learning from “negative examples” facilitates both classification and segmentation of classes of interest. This sheds new light on the computer-aided diagnosis of rare diseases and emerging pandemics, wherein “positive examples” are hard to collect, yet “negative examples” are relatively easier to assemble. Apart from exceeding prior arts in the ChestXray benchmark, our model is particularly strong in identifying diseases of minority classes, yielding over 3-point improvement on average. Remarkably, when using existing partial labels, our model performance is on-par with that using full labels, eliminating the need for an additional 40% of annotation costs. Code will be made available at https://github.com/MrGiovanni/LabelAssemble.