Object Detection
I am working on residential solar panel detection. I am inexperienced with computer vision and I am having trouble deciding on augmentations. My training set has 7k images that are pretty representative of the data I will be deploying on (aside from the fact they come from a different country with different geographic conditions). How should I go about determining which augmentations would be effective?
Hi @Chad_Brouze - thanks for posting!
Generally, any rotation-based augmentations will be safe (as long as you don’t have directional classes like north-facing-roof).
I suggest starting with no augmentations and then layering in additional augmentations as tests. I would only expect augmentations to add a couple of percentage points of mAP → they should not make or break the model.
Hi @Jacob_Witt, thanks for the feedback. What do you think about colour augmentations such as hue, saturation… for this use case?
My default is to always start slowly with those types of augments, and layer them in as tests once you are getting better-than-OK results.
They will never make a bad model good, but they might give you slightly better handling of tricky edge cases in production.