Pelhan Jer ; Lukežič Alan ; Zavrtanik Vitjan ; Kristan Matej

Abstract

Low-shot counters estimate the number of objects corresponding to a selectedcategory, based on only few or no exemplars annotated in the image. The currentstate-of-the-art estimates the total counts as the sum over the object locationdensity map, but does not provide individual object locations and sizes, whichare crucial for many applications. This is addressed by detection-basedcounters, which, however fall behind in the total count accuracy. Furthermore,both approaches tend to overestimate the counts in the presence of other objectclasses due to many false positives. We propose DAVE, a low-shot counter basedon a detect-and-verify paradigm, that avoids the aforementioned issues by firstgenerating a high-recall detection set and then verifying the detections toidentify and remove the outliers. This jointly increases the recall andprecision, leading to accurate counts. DAVE outperforms the top density-basedcounters by ~20% in the total count MAE, it outperforms the most recentdetection-based counter by ~20% in detection quality and sets a newstate-of-the-art in zero-shot as well as text-prompt-based counting.