YOLOv9 Frigate+: Challenges of cars with covers on them

[ianwood]

Often seeing cars with covers on them being missed or mistaken for USPS or something else despite having tagged hundreds of them in my model. Any suggestions?

[inkeep-ai-assistant[bot]]

Based on the Frigate+ documentation, here are some suggestions for improving detection of covered cars:

Submit More Diverse Training Data
The Frigate+ model training process requires diverse examples to handle edge cases like covered cars (1). You should:

• Submit diverse images - provide at least 100 verified images per camera, including varying conditions like cloudy days, sunny days, dawn, dusk, and night (1)
• Submit both true positives and false positives - aim for 80% true positive submissions and 20% false positives across all images (1)
• Address specific false positives - if experiencing false positives in a specific area, submit true positives for any object type near that area in similar lighting conditions to help teach the model what that area looks like (1)

Adjust Detection Thresholds
Consider lowering your thresholds to capture more borderline detections for training (1). If you have false positives scoring at 68% and true positives at 72%, try lowering your threshold to 65% and submit both true and false positives within that range to help the model learn (1).

Model Training Considerations
When switching to YOLOv9, you may need to adjust your thresholds for some objects (1). The model learns from verified labels, so ensure you're verifying images in Frigate+ - only images with at least one verified label will be used when training your model (1).

Since covered cars represent a challenging edge case that differs significantly from typical car appearances, you may need multiple training cycles and extensive examples before seeing improvement in this specific scenario.

📚 Sources:

1. Models

---

Have another question? Just tag @inkeep.

[NickM-27]

Often seeing cars with covers on them being missed or mistaken for USPS or something else despite having tagged hundreds of them in my model. Any suggestions?

Increase the min_score for USPS, probably to 0.8

[ianwood]

OK. Was at 0.65 before. Will try 0.8.

[ianwood]

Now it's getting a false USPS on an uncovered car:

Is there any way to look at my specific Frigate + model to see all the tags associated with each label? Might be a really helpful way to double check that I haven't made any mistakes.

I am 99.9% certain I have not tagged anything incorrectly for USPS but something seems to be odd.

[NickM-27]

I haven't seen this using 0.8. You may just need to do some more training on that to say it is not usps

[ianwood]

I have 32,241 cars tagged in 3,720 images in my Frigate+ model so hopefully not training it what isn't USPS.
Could it be the inverse? I only have 22 USPS tagged so far.

[NickM-27]

yes, training on USPS logos will help it as well.

I would suggest looking at the explore view and seeing what the score of USPS is for these objects

[ianwood]

It seems to be flagging any car as USPS. It's interesting that it's only ever one car at a time. And I wonder if the tracked objects get swapped with a parked car when an actual USPS goes by.