Towards Open-World Marine Intelligence: From Instance Perception to Video Understanding

PhD Qualifying Examination


Title: "Towards Open-World Marine Intelligence: From Instance Perception to
Video Understanding"

by

Mr. Yiwei CHEN


Abstract:

Open-world marine visual understanding is challenging due to underwater visual
degradation, specialized domain knowledge, and the gap between general visual
representations and hierarchical marine taxonomies. While recent marine datasets
and biological foundation models such as BioCLIP have expanded recognition
scale, whole-image recognition remains insufficient in ocean habitats like coral
reefs, where gregarious and cohabiting species require discriminating individual
instances. In this survey, we comprehensively review the marine domain
challenge, and then the evolution of marine instance perception alongside
vision-language models, from closed-set baselines to both open-vocabulary
instance grounding and autoregressive instance-level recognition. Beyond static
images, we examine the progression of instance understanding, tracking, and
individual counting in crowded marine videos. Finally, we discuss the future
direction on open-world marine video understanding with flexible instance
retrieving and grounding across long, untrimmed survey videos.


Date:                   Friday, 20 November 2026

Time:                   3:00pm - 5:00pm

Venue:                  Room 3494
                        Lifts 25/26

Committee Members:      Prof. Sai-Kit Yeung (Supervisor)
                        Prof. Pedro Sander (Chairperson)
                        Prof. Chi-Keung Tang