In practical document processing scenarios, datasets typically grow over time, and their class annotations undergo continuous refinement. This results in significant re-annotation efforts, which are time-consuming and costly.
A promising remedy is to manually re-annotate only a small subset of available documents and apply semi-supervised learning techniques that leverage both labeled and unlabeled data. While numerous approaches exist for tackling classification issues, no adaptation has been made for re-classifying object detection instances, particularly in document layout analysis.
To address this, we propose Bounding Box Label Propagation (BBLP), a pseudo-labeling framework for object detection. An object encoder integrates visual, textual, and positional embeddings from object detection samples to create a joint embedding that can be utilized for Label Propagation on partially annotated datasets in a plug-and-play fashion.
Evaluation results indicate that the proposed approach yields high-quality class annotations of bounding boxes. In the D4LA layout analysis dataset, it achieves a mAP of 54.0%, corresponding to 81.6% of fully supervised performance, using only 10% labeled data. Our work demonstrates the potential of Label Propagation for object detection and lays the groundwork for reducing manual annotation efforts in real-world document processing applications.
Blogger's Review: The BBLP method introduced in this paper effectively enhances the annotation accuracy of object detection by integrating multiple information sources, particularly advantageous when labeling resources are limited. Future research could explore optimizing pseudo-label generation strategies to improve model robustness and precision.