Robust Training of Object Detection with Sparse Annotations

It is difficult to annotate all instances for object detection tasks in practice, especially facing the crowned scenes. The model performance could be impaired when training on sparsely annotated datasets, compared to fully annotated ones. To alleviate this problem, we propose a novel robust training framework for sparsely annotated object detection. We take the sparse annotations as a special case of noisy annotations, where the instances without annotations are regarded as background. First, we propose an Ensemble Prediction (EP) module from different temporal models to avoid the model fitting to the background (noisy labels). Further, an Ensemble Models Non-Maximum Suppression (EM-NMS) module is proposed based on the EP module to generate more robust bounding boxes. Finally, the Pseudo Labels Generating (PLG) module is used to produce pseudo labels for the possible missing-annotation instances. Extensive validation experiments are conducted on PASCAL VOC and MS COCO datasets with sparse annotations. The results show that our proposed robust training framework can further improve model performance when compared to related studies.