The paper proposes a new benchmark for open-world object detection (OWOD) where novel classes are only encountered during the inference stage, unlike previous benchmarks where some novel classes appear in the training set.
The authors introduce a YOLO-based detector called YOLOOC that uses label smoothing to prevent the model from over-confidently mapping novel classes to known classes. The label smoothing technique makes the model less confident about the features it has learned, allowing it to better separate known and novel classes.
Extensive experiments on the new benchmark demonstrate the effectiveness of YOLOOC in discovering novel classes compared to other state-of-the-art OWOD detectors. The paper also conducts an ablation study to analyze the impact of the label smoothing technique and the scaling parameter.
The key highlights of the paper are:
לשפה אחרת
מתוכן המקור
arxiv.org
תובנות מפתח מזוקקות מ:
by Qian Wan,Xia... ב- arxiv.org 04-02-2024
https://arxiv.org/pdf/2404.00257.pdfשאלות מעמיקות