Open-vocabulary object detection via debiased curriculum self-training

Hanlue Zhang, Dayan Guan, Xiangrui Ke, Abdulmotaleb El-Saddik, Shijian Lu

Abstract: Highlights•Open-vocabulary object detection without using box-annotated images of novel classes.•Better exploitation of image-level weak supervision for novel class training.•Proposed debiased curriculum self-training for accurate pseudo-label generation.•Achieved superior performance over two open-vocabulary detection benchmarks.