We presented two papers at the 28th ACM International Conference on Multimodal Interaction (ICMI 2026) in Naples: a full paper titled “Local Optical Flow for Eye Movement Event Detection in Head-Mounted Setups” and a demo titled “ConceptLens: Interactive Concept Bottlenecks for Black-Box Models”.
ICMI is one of the main venues for research on multimodal interaction, and only 36% of regular papers were accepted this year. The full paper addresses eye movement event detection in head-mounted eye tracking, where many algorithms struggle with smooth pursuit; we replaced global scene motion with local optical flow around the gaze point, evaluated the method on a public dataset and showed that it improves smooth pursuit detection over established baselines while running in real time.
To help users act on explanations rather than only read them, our demo presents ConceptLens, a model-agnostic framework for probing pre-trained classifiers; users explore model decisions through semantic concepts instead of individual pixels or tokens and can refine these concepts through feedback; we demonstrated it on image and text datasets from different domains. The code is on GitHub via this link.
Abdulrahman M. Selim from IML at the ICMI conference
