Sensitivity Effects Near Grid Boundaries

[viplix3]

Hi @glenn-jocher,

You opened an issue regarding sensitivity near grid boundaries on AlexyAB's fork of darknet here: https://github.com/AlexeyAB/darknet/issues/3293

I am trying to train a YOLOv3 based object detection model and facing a similar issue. I've gone through the whole thread but couldn't find any conclusive outcomes.

I've posted some follow up questions on that thread. https://github.com/AlexeyAB/darknet/issues/3293#issuecomment-664154657

I was wondering if you were able to solve the said issue.

[github-actions[bot]]

Hello @viplix3, thank you for your interest in our work! Please visit our Custom Training Tutorial to get started, and see our Jupyter Notebook , Docker Image, and Google Cloud Quickstart Guide for example environments.

If this is a bug report, please provide screenshots and minimum viable code to reproduce your issue, otherwise we can not help you.

If this is a custom model or data training question, please note that Ultralytics does not provide free personal support. As a leader in vision ML and AI, we do offer professional consulting, from simple expert advice up to delivery of fully customized, end-to-end production solutions for our clients, such as:

• Cloud-based AI systems operating on hundreds of HD video streams in realtime.
• Edge AI integrated into custom iOS and Android apps for realtime 30 FPS video inference.
• Custom data training, hyperparameter evolution, and model exportation to any destination.

For more information please visit https://www.ultralytics.com.

[glenn-jocher]

@viplix3 yes. I discovered this issue a while back after discovering hotspots in pedestrian detection using our iOS app. The issue is resolved in YOLOv5. If you see evidence to the contrary though please let us know.

Single Frame	2D Histogram over 1 hour

[viplix3]

@glenn-jocher Thanks for the clarification. Closing.

[Zubair2019]

Surely this problem has been addressed by the authors of YOLO(v4) by calculating the dimensions of bounding box differently.They solved this problem through multiplying the sigmoid by a factor exceeding 1.0, so eliminating the effect of grid on which the object is undetectable.

[glenn-jocher]

@Zubair2019 eliminating is a strong word for this. Reduced is more realistic.

I've also never repeated the above experiment, so no fix has ever been verified or observed directly. It's likely all xywh spaces will suffer from this problem to varying degrees.