At BCI 2026 (12th Balkan Conference on Informatics) in Thessaloniki, Iason Karakostas from CERTH presented the work: “Pointing-attention guidance for improved few-shot object detection”.

Vision-Language Models (VLMs) like CLIP excel at open-vocabulary tasks but treat entire

images uniformly, often struggling to localize objects when training data is scarce.

Drawing inspiration from how humans use spatial context, the work integrates spatial

priors, such as pointing cues encoded as Gaussian heatmaps, directly into the training

objective of region-based VLM detectors.

Pointing guidance is strictly applied during training; the fine-tuned detector operates on

standard RGB images at test time without requiring human cues.

Evaluated on a challenging few-shot setup, the method delivered a +4.34 AP overall gain

and a notable ~+10 point increase at AP75, significantly sharpening bounding box

accuracy in cluttered environments.