Consistency of performance gains with increasing data scale for label-efficient detectors

Determine whether label-efficient object detection paradigms—including label-noise–robust learning, semi-supervised learning, weakly supervised learning, sparse-shot learning, and zero-shot detection—exhibit consistent performance improvements as dataset size increases.

Background

Label-efficient object detection methods aim to reduce annotation costs through alternative supervision regimes such as robustness to label noise, semi-supervised learning, weak supervision, sparse-shot, and zero-shot detection. While widely validated on generic object detection, their behavior as the amount of training data scales up has not been clearly characterized.

The paper introduces TinySet-9M and a benchmark to study small object detection under diverse supervision regimes, motivating the need to understand whether these label-efficient approaches benefit consistently from larger datasets.

References

Although these approaches have been extensively validated and widely adopted in generic object detection, their effectiveness on small object detection has not yet been systematically analyzed. In parallel, it remains unclear whether these algorithms exhibit consistent performance gains as data scale increases.

— Generalized Small Object Detection:A Point-Prompted Paradigm and Benchmark  (2604.02773 - Zhu et al., 3 Apr 2026) in Related Works, Label-efficient Object Detection subsection