2 Eyes and 1 Mouth (MSc Thesis)
Agentic Image Segmentation and Object Video Tracking for Cross-View Correspondence.
Task: Cross-view Object Correspondence — given two synchronised videos of the same scene and a target object, match the mask of the object for all frames.
Dataset: Ego-Exo4D.
Challenge: videos are taken from different perspectives, therefore the target object changes appearance, size, position in the frame, and possibly gets occluded or leaves one of the two videos entirely.
Gap in the field: all models rely on dense visual representations and trained architectures, which makes them uninterpretable by design and arguably limited to the distribution seen at training.
My thesis: a training-free pipeline that leverages natural language descriptions of the target object to bridge the two views.
Result: graduation grade of 110/110 (first class honours).
