ABC Easy as 123: A Blind Counter for Exemplar-Free Multi-Class Class-agnostic Counting

TL;DR

ABC123 introduces an exemplar-free, multi-class class-agnostic counting framework built on density-map regression with a multi-head transformer backbone, paired with a Hungarian-based matching stage and an interpretability-focused example-discovery module. The authors also present MCAC, a synthetic dataset with 1–4 object classes per image and detailed annotations to study counting across multiple unseen classes. Empirical results show ABC123 surpasses exemplar-based counting methods on MCAC and transfers effectively to FSC-147/133, while offering explanations of counted objects through discovered examples. This work advances practical, scalable counting in multi-class, open-set scenarios and provides a new benchmark to evaluate exemplar-free counting approaches.

Abstract

Class-agnostic counting methods enumerate objects of an arbitrary class, providing tremendous utility in many fields. Prior works have limited usefulness as they require either a set of examples of the type to be counted or that the query image contains only a single type of object. A significant factor in these shortcomings is the lack of a dataset to properly address counting in settings with more than one kind of object present. To address these issues, we propose the first Multi-class, Class-Agnostic Counting dataset (MCAC) and A Blind Counter (ABC123), a method that can count multiple types of objects simultaneously without using examples of type during training or inference. ABC123 introduces a new paradigm where instead of requiring exemplars to guide the enumeration, examples are found after the counting stage to help a user understand the generated outputs. We show that ABC123 outperforms contemporary methods on MCAC without needing human in-the-loop annotations. We also show that this performance transfers to FSC-147, the standard class-agnostic counting dataset. MCAC is available at MCAC.active.vision and ABC123 is available at ABC123.active.vision.

Figures (8)

• Figure 1: ABC123 counts objects of multiple unseen types. Not only does our method not need exemplars to define the type to count, it finds examples of each type it has counted.
• Figure 2: MCAC Contains images with up to 4 classes and up to 300 instances per class. All objects have associated instance labels, class labels, bounding boxes, centre points, and occlusion percentages.
• Figure 3: The ABC123 pipeline. Our method learns to count objects of multiple novel classes without needing exemplar images. During training and quantitative evaluation, the matcher aligns the unguided predictions to the ground truth labels. To aid a user in understanding the results, the example prediction stage locates instances associated with each generated count.
• Figure 4: Example Finding. Instead of using exemplars to define the count, we count 'blind' and then find meaningful bounding boxes to aids a user in understanding what has been counted. The examples are found using the query image, learnt features, our regressed density maps and a SAM network.
• Figure 5: Comparison to other methods on MCAC. ABC123 produces more accurate results than the exemplar-based methods without using exemplar images. The ground truth (GT) and predicted counts are shown in the top right corner of their respective density maps.