EE Seminar: Is There a Needle in the Haystack
Electrical Engineering Systems Seminar
Speaker: Shoval Mishal
M.Sc. student under the supervision of Prof. Shai Avidan
Wednesday, 13th November 2024, at 15:00
Room 011, Kitot Building, Faculty of Engineering
Is There a Needle in the Haystack?
Abstract
We are interested in detecting the existence of novel object classes in aerial images, without specifying these novel classes ahead of time.
In our setting, we are equipped with a detector capable of detecting a closed set of objects (e.g., vehicles, planes) but wish to determine if other, unspecified, object classes, that are of interest (say, ships), appear in the images as well. This open vocabulary problem poses two challenges. The first is scale, as there are tens of millions of patches to evaluate. The second is vagueness. How do you determine whether a given patch contains a semantically meaningful object of interest or just an interesting background pattern?
To address these challenges, we propose a funnel approach that gradually reduces the number of patches of interest from tens of millions to a short list of few tens of thousands. The patches in the short list are ranked automatically and shown to a human operator. We therefore measure performance by ``Time-To-$1^{st}$'' (TT-1), i.e. the time it takes a human to find the first instance of interesting new classes in aerial images, and show we are capable of producing such a sample within the first few patches. Our code will be made publicly available.
השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום שם מלא + מספר ת.ז. בדף הנוכחות שיועבר באולם במהלך הסמינר
