Bounding Box Label Propagation for Re-Annotation of Document Layout Analysis Datasets

The article presents Bounding Box Label Propagation (BBLP), a novel pseudo-labeling framework designed for re-annotating object detection instances in document layout analysis. By integrating visual, textual, and positional embeddings, BBLP achieves a mean Average Precision (mAP) of 54.0% on the D4LA dataset using only 10% of labeled data, which is 81.6% of the performance of fully supervised methods. This approach significantly reduces the manual effort required for dataset re-annotation, making it a valuable tool for practitioners in document processing who seek efficient ways to maintain and improve their annotation quality.