RMSOD: A RetinaNet based Mixed Supervised Object Detection Framework
Blossom Treesa Bastian, C. V. Jiji · 2022 IEEE 19th India Council International Conference (INDICON) · 2022
Mixed supervised object detection (MSOD) exploits a source dataset with full instance-level annotations and a target dataset with few instance-level annotations to detect new categories in the target dataset. In this work, we propose a RetinaNet based MSOD by employing a three-stage training pipeline, where the detector is trained on a source dataset with complete bounding box information, a target dataset with image-level annotation and on few images from target dataset with instance-level annotation. The proposed RMSOD framework effectively makes use of the readily available image-level annotation of the target dataset, but do not rely on any kind of relationship between object categories in source and target datasets unlike existing MSOD frameworks. The mAP results on PASCAL VOC 2007 test dataset shows the efficacy of the proposed MSOD configuration over existing few-shot object detection algorithms based on weakly-supervised, semi-supervised and mixed-supervised learning. Furthermore, by employing RetinaNet as baseline detector we ensure the deployability of the proposed detector for real-time implementation.