Online Pseudo-Label Unified Object Detection for Multiple Datasets Training

TL;DR

This paper conducts a thorough analysis of the cross datasets missing annotations issue, and proposes an Online Pseudo-Label Unified Object Detection scheme that uses a periodically updated teacher model to generate pseudo-labels for the unlabelled objects in each sub-dataset.

Abstract

The Unified Object Detection (UOD) task aims to achieve object detection of all merged categories through training on multiple datasets, and is of great significance in comprehensive object detection scenarios. In this paper, we conduct a thorough analysis of the cross datasets missing annotations issue, and propose an Online Pseudo-Label Unified Object Detection scheme. Our method uses a periodically updated teacher model to generate pseudo-labels for the unlabelled objects in each sub-dataset. This periodical update strategy could better ensure that the accuracy of the teacher model reaches the local maxima and maximized the quality of pseudo-labels. In addition, we survey the influence of overlapped region proposals on the accuracy of box regression. We propose a category specific box regression and a pseudo-label RPN head to improve the recall rate of the Region Proposal Network (PRN). Our experimental results on common used benchmarks (\eg COCO, Object365 and OpenImages) indicates that our online pseudo-label UOD method achieves higher accuracy than existing SOTA methods.

Figures (9)

• Figure 1: (a) Ambiguous background problem: some categories in one dataset are not annotated in other datasets. (b) Our proposed Online Pseudo-Label UOD (OPL-UOD) uses a periodically updated teacher model to generate pseudo-labels of the unlabelled objects in the datasets, which enables the model training to obtain more cross datasets annotations and improved the UOD performance. Blue arrows indicate the manual label supervision training process, and red arrows indicate the pseudo-label supervision training process.
• Figure 2: The overlapped boxes problem. The top row demonstrates that the annotation boxes of certain classes are prone to overlap. The bottom row shows the positive region proposals, which are marked with the corresponding colors to represent different ground truth categories.
• Figure 3: The baseline UOD structure.
• Figure 4: The top figure indicates that both the offline pseudo-label training and the EMA online pseudo-label UOD uses a step learning rate schedule (yellow curve), and the proposed online pseudo-label UOD used a cosine learning rate schedule (green curve). The bottom figure shows the mAP scores on the COCO-Split5 datasets. After adopting the periodical teacher updating, the mAP score of the cycle student model reaches the local maxima at minimum points of the cosine learning rate. The mAP score of the cycle teacher model is much higher than the EMA teacher.
• Figure 5: Pseudo-labels of different EMA teacher models on the COCO, Object365 and OpenImages datasets.