Drone detection experiment pipeline
A recoverable experiment system for comparing detection models across heterogeneous datasets—not merely a single training run.
Data engineering
One task, incompatible sourcesThree visible-light drone datasets arrived as VOC XML and video/XML combinations. They were normalized into YOLO and COCO representations with consistent class semantics and dataset splits.
Pipeline acceleration
Batch the expensive boundaryThe original ARD-MAV process invoked ffmpeg frame by frame and was projected to take about 18 days. Exporting all frames once per video reduced the process to roughly 10–20 minutes.
Experiment operations
Make failures recoverableThe system included environment checks, small-subset smoke tests, multi-GPU scheduling, manifests, checkpoint reuse, three-seed tracking, and lightweight result aggregation.
Claim boundary
Design versus completed runsThe complete design covered six model families and three seeds. The strongest fully completed core comparison included RT-DETR-L and Faster R-CNN across three seeds each; other model results are reported separately with their own stability caveats.