Luca Ciampi: How machines count – Visual object counting from limited supervision to the open world

Counting objects in images and videos is a deceptively simple task. Modern deep-learning approaches can achieve impressive accuracy, yet they often rely on large, carefully annotated datasets and struggle with new environments, unseen camera viewpoints, or previously unseen object categories. Even recent vision-language models, despite their broad recognition capabilities, remain unreliable at counting, particularly in crowded scenes. This seminar examines visual object counting when annotated data are scarce or costly, labels vary across annotators, or supervision is unavailable for new domains or object categories.

The talk will begin with class-specific counting in images and videos, showing how synthetic data can reduce annotation costs and how unsupervised domain adaptation can bridge the gap between training and deployment conditions. It will then turn to settings in which annotations are available but uncertain or inconsistent. Using biological structure counting as a case study, it will examine how disagreement among multiple human raters can be modeled rather than collapsed into a single consensus ground truth. The final part will focus on class-agnostic counting, where the target object class is specified at inference time through visual exemplars or natural-language prompts. For exemplar-based counting, the talk will explore how self-supervised visual representations can support training-free methods without annotated training data. For text-guided counting, it will examine the use of vision-language models and the limitations of current evaluation protocols. Conventional counting metrics can conceal failures of semantic grounding: a low counting error does not necessarily mean that a model has understood which objects the prompt refers to.

Across applications ranging from traffic and crowd monitoring to biomedical imaging, the seminar will highlight how different forms of limited supervision call for different solutions, and will discuss the remaining challenges in building counting systems that are scalable, semantically grounded, and reliable in the open world.

Dr Luca Ciampi is a researcher in the Artificial Intelligence for Media and Humanities Laboratory at ISTI-CNR in Pisa. He received his PhD in Information Engineering from the University of Pisa, with a thesis on deep-learning techniques for visual counting. His research explores visual and multimedia understanding when labelled data are scarce, noisy, or costly to obtain. His main focus is object counting and localisation in images and videos, spanning domain adaptation, synthetic data, multi-rater learning, and class-agnostic approaches. He also works on biomedical imaging and bio-inspired learning, including biological neural cultures as computational reservoirs. His work has received two paper awards; he was named a Key Innovator by the European Commission, and his MSCA proposal received a Seal of Excellence.