search

Yang7879 / 3D-BoNet

🔥3D-BoNet in Tensorflow (NeurIPS 2019, Spotlight)
https://arxiv.org/abs/1906.01140
MIT License
392 stars 85 forks source link

Question on S3DIS #37

Open LeeLinSCUT opened 4 years ago

LeeLinSCUT commented 4 years ago

Thx for your working. I've trained the module with 50 epochs and finnaly finish the evalution on Area5. But the results are quiet different from your paper table 3. The Area 5 mPrec and mRecall are 57.5 and 40.2 but the mPrec I got is 52%. And then I use the realeased module. The mPrec is 60 which is close enogh. I'm wondering if the released module was trained on different hyperparameters or is there other reasons?

Yang7879 commented 4 years ago

Hi @LeeLinSCUT, the hyperparameters are the same as what I used to train, but the results may indeed be different (can be higher or lower), probably because the association algorithm is not always stable enough.