Reproduce “Unlocking ImageNet's Multi-Object Nature: Automated Large-Scale Multilabel Annotation” on OpenResearch

About this paper

Junyu Chen, Md Yousuf Harun, Christopher Kanan

Researchers from the University of Rochester and Rochester Institute of Technology developed a fully automated pipeline to convert the ImageNet-1K training set into an explicit, region-grounded multi-label dataset, addressing its long-standing single-label limitation. Training with these new annotations improved ResNet-50 top-1 accuracy by up to 2.0% on ImageNet-ReaL and 1.5% on ImageNet-V2, and enhanced transfer performance on downstream multi-label tasks by up to 2.0% mAP on COCO.

Key insights

• An automated, region-grounded multi-label annotation pipeline can effectively overcome the single-label limitation of large datasets like ImageNet without requiring human intervention.
• Training with rich multi-label supervision leads to more robust and generalizable visual representations, yielding higher classification accuracy and better transferability compared to traditional single-label training.
• The 'Local-Soft + Original' label aggregation strategy, combining soft per-mask probabilities with the original single ImageNet label, proved most effective for deriving image-level multi-labels.

Summary by alphaXiv.