I wish to be able to detect: pedestrians, cars, traffic lights

I have two large datasets: - One contains instances and labels of all three classes. - The other contains instances of all three but only labels for pedestrians and cars. ie. there are many unlabelled traffic lights.

I want to combine the two datasets and train Yolov3 on it. Will the unlabelled presence of objects of interest significantly affect detection performance of that category?


So depends on sizes— if the second dataset has a lot of traffic lights without labels and you in a basic setup train on it, it will cause large performance decreases on that class. If it’s small, then it may just be “noise” to the model and it’ll filter it out, and still perform well. My recommendation is to adjust the network slightly and add a flag input saying which dataset the input was from: if from first, train normally. if from the second then ignore all losses pertain to boxes/classes that the network thinks a traffic light was in: that way your labels don’t try to tell the mode l “there’s no traffic light there”

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.