Multi-class density estimators are commonly used to count objects in crowded scenes, but their class-wise outputs say nothing about how reliably the density mass has been assigned. Without a confidence signal, it is hard to know whether a particular count is trustworthy, especially when classes are visually similar or overlap.
The paper introduces a training-free confidence proxy built on local class competition. The idea is that when multiple classes compete for the same density mass, the strength of that competition can indicate whether the allocation is dependable. Because the proxy is computed from the estimator's own outputs, it adds no training overhead and requires no extra labels.
This reframes confidence as a byproduct of the counting process itself, addressing a practical gap in density-based counting: not just how many objects are present, but how sure the model is about what it is counting. The abstract does not include experimental results, so the real-world impact of the proxy remains to be demonstrated.