Iterative weakly supervised learning for novel class object detection

Dejana Mandić, Wieland Brendel, Claudio Michaelis

Keywords: object detection, novel class, weakly supervised

TL;DR: A framework for training novel-class object detectors using a combination of a few labeled images and weakly labeled data.

Abstract: Training object detectors for new classes usually requires collecting and labeling large amounts of data. Our paper introduces a new approach to address this issue - training novel-class object detectors using a combination of a few labeled images and weakly labeled data, that is easy to obtain. We propose an iterative fine-tuning framework that cycles through predicting pseudo-labels, filtering them using weak labels, and fine-tuning the model on this data. By repeating the process, we can mostly close the gap to a model trained on 40x more data, thereby offering a new approach to improving the trade-off between labeling effort and performance.