Structured Ablation of YOLOv8-L on VisDrone
Four modifications from the UAV small-object detection literature — attention, a stride-4 detection head, PIoU regression reweighting, and inference-time tiling — applied to one YOLOv8-L baseline under a single protocol. None beat the unmodified model. The most-recommended fix cost 0.148 mAP50 and reduced small-object recall by 31%.
CASE STUDY
Project breakdown.
Select a section to explore the project one part at a time.
Problem
Every published modification claims a gain. This one tested four.
UAV imagery breaks object detectors in a specific way. Objects occupy a handful of pixels, cluster densely, and occlude each other constantly. The published response is a standard toolkit — attention modules, an extra high-resolution head, a reweighted regression loss, inference-time tiling — and papers reporting these almost always report gains. This project applied all four to one baseline, under one protocol, and measured each independently.
What this project asks
- Does attention help where clutter is the problem? EMA-style attention inside C2f blocks is meant to sharpen feature selectivity in exactly these scenes.
- Does a higher-resolution head recover small objects? A stride-4 P2 head preserves shallow spatial detail that deep backbones destroy. The reasoning is sound. Is the result?
- Can loss shaping fix what architecture broke? PIoU-inspired regression reweighting targets the instability of tiny-box regression directly.
- Is inference-time tiling universally useful? SAHI slices the image so small objects occupy more pixels per tile. Does that hold across every model it is applied to?
VisDrone2019-DET, 10 classes, 38,759 annotated instances across the 548-image validation split. YOLOv8-L at 640×640, 150 epochs, batch 16, on an NVIDIA A100. Every stage inherits the same data, schedule, and evaluation code.