Hi everyone,
I'm working on a cotton detection project using OAK-D Lite with a YOLOv11 segmentation model. I'm facing a challenging issue with specular reflections and would appreciate your guidance.
Problem:
My YOLOv11 model detects cotton bolls accurately in most cases
However, specular highlights/reflections on leaf surfaces are being falsely detected as cotton
These reflections have similar color (white/bright) and shape characteristics to actual cotton
This happens with both bright spots and sometimes entire leaf surfaces appearing cotton-like
Current Setup:
Camera: OAK-D Lite
Model: YOLOv11 segmentation
Environment: Outdoor agricultural field (natural sunlight)
Using either HFoV or VFoV mode
What I've Tried:
Post-processing filters (color, texture, edge analysis)
Confidence thresholds
Shape analysis on segmentation masks
None of these software approaches work reliably because the reflections genuinely look like cotton in the captured images.
Questions:
Does Luxonis offer or recommend a polarizing filter compatible with OAK-D Lite?
- Circular polarizer (CPL) to reduce specular reflections?
- Any specific models/brands that fit the OAK-D Lite lens?
What are the lens specifications for OAK-D Lite?
Alternative hardware solutions?
Is there a way to adjust camera settings (exposure, gain, etc.) to minimize reflections?
Would a lens hood or shade help?
Has anyone solved similar reflection issues with OAK-D cameras?
Additional Context:
I prefer a hardware solution over retraining the model, as it would be cleaner and more reliable. However, if no polarizing filter option exists, I'll proceed with collecting hard negative examples and retraining.
Any guidance, part numbers, or experience with this would be greatly appreciated!
Thanks in advance!