A Method for Low-shot Object Counting and Detection

Pelhan, Jer (author), Kristan, Matej (mentor)

We tackle the problem of few-shot object counting and detection of arbitrary object categories using only a small number of annotated instances, namely exemplars. The task of the method is to count and detect all objects that are part of the same semantic category as the exemplars but may vary widely in visual appearance. Methods currently address this problem by generalizing the appearance of the exemplar objects, allowing them to count effectively. The generalization capacity leads to high recall, but also to low precision, due to non-discriminative counting. Few-shot counting methods predict solely the total count of objects and do not provide estimations of their locations, which is crucial with many applications. We propose a novel method DAVE (Detect and Verify), which aims to bridge the gap between traditional few-shot counting methods and the emerging field of few-shot counting and detection, by predicting accurate count and locations of objects. We introduce a detect-and-verify paradigm into few-shot counting, achieving both high recall and precision rates. DAVE outperforms the most recent detection counter by 20% in terms of AP50 and decreases the counting error of the top-performing counter by 20% in terms of MAE. DAVE further outperforms all 0-shot counters in terms of RMSE, and achieves on-par AP50 performance as the best few-shot counting and detection method.

• Language: English
• Keywords: computer vision, object counting, object detection, few-shot learning
• Work type: Master's thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2023
• PID: 20.500.12556/RUL-149978
• COBISS.SI-ID: 165352707
• Publication date in RUL: 12.09.2023

Secondary language

• Language: Slovenian
• Title: Metoda za štetje in detekcijo objektov z malo učnimi primeri

Abstract

V tem delu naslovimo problem štetja in detekcije objektov poljubnih kategorij z malo učnimi primeri. Naloga metode je prešteti vse instance objektov, ki so sicer del iste semantične kategorije kot podani primeri, ampak so si vizualno lahko zelo raznoliki. Obstoječe metode omenjeni problem naslovijo z močno generalizacijo izgleda podanih nekaj učnih primerov, kar jim omogoča uspešno štetje. Sposobnost generalizacije sodobnih metod vodi v visoke vrednosti priklica, in znižano natančnost zaradi nediskriminativnega štetja. Poleg tega so metode štetja z malo primeri globalni števci, ki ne zagotavljajo lokacij objektov, kar je ključno pri mnogih aplikacijah. Predlagamo novo metodo DAVE (ang. Detect and Verify), ki poveže obstoječo vrzel med tradicionalnimi metodami za štetje z malo primeri in nastajajočim področjem štetja in zaznavanja z malo primeri, saj omogoča napoved natančnega števila objektov in njihove lokacije. DAVE uspešno uvede paradigmo zaznaj in preveri, ki omogoča doseganje visokih vrednosti priklica in natančnosti. DAVE dosega nižjo napako kot sodobne metode pri nalogi štetja s relativnim izboljšanjem 20% MAE, in pri nalogi zaznavanja za 20% AP50. DAVE doseže manjšo napako RMSE kot sodobni števci, in dosega primerljive detekcijske rezultate v AP50, kot metode, ki na vhod prejmejo primerke.

• Keywords: računalniški vid, štetje objektov, detekcija objektov, učenje z malo učnimi primeri