YOLO False Positives/Negatives in UAV Scenario

Discussion
,
1

Hello everyone,

I’m currently working on a UAV-based object detection project and would really appreciate insights from the community.
I have several years of experience in AI, mainly focused on deep learning and computer vision. Over the past ~1 year, I’ve worked extensively with YOLO-based models across multiple object detection tasks.

Current Challenge

In this UAV scenario, I’m facing persistent issues with both false positives (FP) and false negatives (FN) under challenging conditions:

  • Objects near image boundaries (edges/corners) are often missed or poorly localized
  • Motion blur due to UAV movement significantly degrades detection performance
  • Sensor noise / low-quality frames further reduce object visibility
    Overall, the model struggles to maintain stable precision and recall in real flight conditions.

What I’ve Tried So Far

  • Augmentations: mosaic, translation, scaling, HSV, rotation, etc.
  • Adjusting confidence and IoU thresholds
  • Negative sampling → significantly reduced false positives
  • Experimented with multiple YOLO11-based model variants like MASF-Yolo
  • Adjusted loss weights to emphasize primary objects
  • Tried SAHI (slicing-based inference) for small objects → improved detection in some cases but introduced many additional false positives, making it less usable in my pipeline

Despite these efforts, performance remains inconsistent, especially in edge cases.

Now I 'd like to focus on improving two areas: Dataset, Loss function.

1. What are the most effective dataset strategies to improve robustness for edge cases like objects near image boundaries, motion blur, and noisy frames?
2. Which loss functions or training strategies are best for balancing false positives and false negatives in challenging UAV scenarios?

I know this is challenge but my goal is to develop a model that performs reliably in real UAV conditions, even with degraded image quality, while maintaining strong precision and recall.

Any suggestions, references, or shared experiences would be greatly appreciated.

Thanks in advance!

2

For UAV data, I’d focus much more on data curation than custom loss edits. The biggest wins usually come from explicitly oversampling your hard cases: partially visible border objects, real motion-blurred frames, noisy/compressed frames, and confusing negatives from the same flight domain. Also make sure your labeling policy is consistent for truncated edge objects; inconsistent border labels create both FP and FN very quickly. I’d avoid relying only on synthetic blur/noise and instead mine real bad frames from flight logs, then keep those well represented in train and val.

On the model side, if you’re starting a new round of experiments I’d try Ultralytics YOLO26 since small-object aerial scenarios are exactly where its ProgLoss + STAL improvements help most, as noted in the YOLO26 comparison for drone and aerial imagery and the broader YOLO26 overview. If you stay on YOLO11, I’d prioritize hard-example mining, cleaner border annotations, larger imgsz or controlled tiling, and reducing heavy mosaic late in training over rewriting the loss. SAHI can still work, but usually only after tightening slice fusion/NMS and using stronger negative data to control the FP spike.

If you want, share a few typical FP/FN crops and your class-size distribution and I can suggest a more targeted setup.

3

Hi there, thanks for your reply.

Would it be appropriate to label objects that are difficult to identify due to image blur, even if they may be considered true positives? I believe blur is one of the main factors contributing to the higher false positive rate. The trained model occasionally detects these objects, but I am not fully confident that the detections and classifications are accurate.

Would it be more appropriate to label only well-identified objects and use augmentation to expand the training data?

Thank you