Code, Configurations, and VisDrone2019-DET Train/Validation Data for the LA-DEIM UAV Object Detection Network

This data deposit supports the manuscript “LA-DEIM: A Lightweight and Efficient UAV Detection Network.” It contains the LA-DEIM training, validation, ablation, and knowledge-distillation code; configurations for the baseline, component ablations, full model, and distilled student; and the VisDrone2019-DET training and validation data used in the study.The training split contains 6,471 JPEG images and 353,550 original annotation rows, while the validation split contains 548 JPEG images and 40,169 original annotation rows.

The package provides both the original VisDrone TXT annotations and deterministically generated COCO-format JSON annotations. During conversion, VisDrone categories 1–10 were remapped to contiguous model category IDs 0–9. Ignored regions, categories 0 and 11, records with score 0, and non-positive bounding boxes were excluded.