Abstract
Abstract
Recent years have seen impressive advancements in Object Detection models. How-
ever, these advancements are typically limited to relatively balanced datasets. In a
long-tailed setting, detectors often exhibit a bias towards head classes, resulting in
subpar performance for tail classes. Long-tailed learning is crucial as the distribution
of objects in real life follows the Power Law and Zipf’s Law. Numerous techniques
have been proposed to address this issue. In this thesis, we explore methods from
the most influential branches of long-tailed learning. We then propose two post-hoc
class score calibration methods that utilize training performance measurements, of-
fering an alternative to existing methods that rely on class sample sizes. Furthermore,
we introduce a third method that employs a ranking-based loss function during the
second stage of training. We evaluate these methods using a challenging long-tailed
dataset LVIS and compare our results with recent approaches. Our results demon-
strate that our methods improve upon the baseline established with LVIS and present
competitive performance compared to similar applications.