Akademik Veri Yönetim Sistemi
Tezin Türü: Doktora
Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Mühendislik Fakültesi, Bilgisayar Mühendisliği Bölümü, Türkiye
Tezin Onay Tarihi: 2021
Tezin Dili: İngilizce
Öğrenci: KEMAL ÖKSÜZ
Asıl Danışman (Eş Danışmanlı Tezler İçin): Sinan Kalkan
Eş Danışman: Emre Akbaş
Özet:
This thesis has two aims: (Aim 1) Identifying imbalance problems in visual detection, and (Aim 2) addressing these problems using loss functions based on performance measures. For Aim 1, we present a comprehensive review of the imbalance problems in object detection including a problem-based taxonomy and a detailed discussion for each problem with its solutions and open issues. To achieve Aim 2, we identify two challenges: (i) Average Precision (AP), the common performance measure, has certain drawbacks. To remedy them, we propose Localisation Recall Precision (LRP) Error as a novel performance measure. (ii) Loss functions derived from performance measures are ranking-based functions whose derivatives are zero or infinite, thus, they cannot directly be used with backpropagation. To overcome this, based on perceptron learning, we propose Identity Update, a simple and general optimisation method for ranking-based losses, which provably ensures balance in terms of total gradient mag- nitudes of positives and negatives. Having addressed these challenges, using LRP Error and Identity Update, we propose average LRP Loss and Rank & Sort (RS) Loss for balanced training of visual detectors. We show that our loss functions have the following unique benefits: (i) They are easy-to-tune with a single hyper-parameter, different from common methods with ~7 hyper-parameters on average, (ii) they en- force correlation among sub-tasks of visual detectors (i.e. classification and different localisation tasks), which affects both the remaining detections after Non-Maximum- Suppression and performance measure AP, and (iii) they are applicable to a diverse set of visual detectors (i.e. one-stage, multi-stage, anchor-based, anchor-free, with balanced or severely imbalanced data). As a result of these benefits, for example with RS Loss, we train four object detection and three instance segmentation methods only by tuning the learning rate and consistently improve their performance.