Calibrated Teacher for Sparsely Annotated Object Detection

Fully supervised object detection requires training images in which all instances are annotated. This is actually impractical due to the high labor and time costs unavoidable missing annotations. As a result, incomplete annotation each image could provide misleading supervision harm training. Recent works on sparsely annotated alleviate this problem by generating pseudo labels for Such mechanism sensitive threshold of label score. However, effective different stages among detectors. Therefore, current methods with fixed thresholds have sub-optimal performance, difficult be applied other In order resolve obstacle, we propose Calibrated Teacher, confidence estimation prediction well calibrated match its real precision. way, detectors would share similar distribution output confidence, so that multiple same achieve better performance. Furthermore, present simple but Focal IoU Weight (FIoU) classification loss. FIoU aims at reducing loss weight false negative samples caused annotation, thus as complement teacher-student paradigm. Extensive experiments show our set new state-of-the-art under sparse settings COCO. Code will available https://github.com/Whileherham/CalibratedTeacher.