Reproduce “Object Detection as a Positive-Unlabeled Problem” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuewei Yang, Kevin J Liang, Lawrence Carin
A framework recasts object detection as a Positive-Unlabeled (PU) learning problem to manage incomplete dataset annotations. This approach consistently enhances model training robustness against missing ground truth labels and yields performance improvements across both artificially corrupted and real-world datasets.
Key insights
- Object detection, given practical annotation challenges, is more accurately framed as a Positive-Unlabeled problem than a Positive-Negative one.
- Allowing models to classify unlabeled regions as positive without penalty provides a more appropriate learning signal for datasets with incomplete annotations.
- Dynamically estimating the positive class prior within the PU framework is crucial for adapting to the evolving nature of object proposals during detector training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.