Reproduce “WUTDet: A 100K-Scale Ship Detection Dataset and Benchmarks with Dense Small Objects” on OpenResearch

About this paper

Junxiong Liang, Mengwei Bao, Tianxiang Wang, Xinggang Wang, An-An Liu, Ryan Wen Liu

Researchers from Wuhan University of Technology and collaborators introduced WUTDet, a large-scale ship detection dataset comprising over 100,000 images and nearly 400,000 annotated ship instances, designed to feature dense small objects and diverse maritime conditions. The study provides comprehensive benchmarks for 20 mainstream object detection models across CNN, Transformer, and Mamba architectures, demonstrating that models trained on WUTDet achieve superior generalization performance on a unified cross-dataset test set (Ship-GEN).

Key insights

• The WUTDet dataset's scale and diversity, particularly its rich representation of small objects and challenging environmental conditions, enable models to learn more robust and generalizable features for ship detection.
• Transformer-based models consistently achieve the highest detection accuracy on WUTDet, especially for small objects and in adverse conditions like fog and low-light, despite often having lower inference speeds compared to CNNs.
• Training models on WUTDet significantly improves their generalization performance on unseen, mixed-source data (Ship-GEN), indicating the dataset's effectiveness in fostering models less sensitive to data distribution shifts.

Summary by alphaXiv.