search

Zhongdao / Towards-Realtime-MOT

Joint Detection and Embedding for fast multi-object tracking
MIT License
2.37k stars 539 forks source link

Question about Automatic Loss Balancing Function #172

Closed roborocklsm closed 3 years ago

roborocklsm commented 4 years ago

Hey, thanks for your great work! But I have a question about the Automatic Loss Balancing Function.

According to Kendall, Gal, andCipolla 2018, continuous output and softmax output have different coefficients: continuous one contains 1/2 while softmax does not. But I notice both in your paper and code give coefficient 1/2 to all three losses.

Would you mind giving some further explanations?

Zhongdao commented 3 years ago

Yes, there should be a 1/2 in softmax according to the derivation. But in practice, we find this constant factor hardly influence the final performance, so we remove the 1/2.

roborocklsm commented 3 years ago

@Zhongdao Yes, sure, you are right. I have tried and it does have no difference between two results.

Thx.