Giters

kkhoot / PAA

A PyTorch implementation of the paper `Probabilistic Anchor Assignment with IoU Prediction for Object Detection` ECCV 2020 (https://arxiv.org/abs/2007.08103)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Question about the hyper parameters of GMM

Chrisfsj2051 opened this issue · comments

commented

Hi @kkhoot ,

Thanks for your amazing and stable work!

I'm trying to implement PAA, but I'm confused about the argument precisions_init. What's the reason for not using the default value of this argument?

PAA/paa_core/modeling/rpn/paa/loss.py

Line 193 in 65bde62

precisions_init=[[[1.0]], [[1.0]]]

commented

Hi, thanks for your interest!
The default init methods may also work well. The init mean/precision values in our paper are to make anchor scores are well separated with few (45) examples. If you perform EM with all the anchors, then I think it should be less sensitive with init methods or values. However, because of the speed issue, only 45 samples are collected for EM. So I set the mean values as min/max scores to make them easily separated. The precision values (1.0, 1.0) are empirically set, and I guess the exact value should not be important and other values can also work well.

commented

Thanks!