search

heng-hw / SpaCap3D

[IJCAI 2022] Spatiality-guided Transformer for 3D Dense Captioning on Point Clouds (official pytorch implementation)
https://spacap3d.github.io/
20 stars 5 forks source link

About the loss curve #1

Closed pilot2022 closed 2 years ago

pilot2022 commented 2 years ago

Hello, thanks for your work and code!

Why the loss curves of x and y are unstable while z is stable？

0d2c596b6092d1bfc55ddb447a42399

Did this happen during your training?

heng-hw commented 2 years ago

Hello @pilot2022,

Thanks for your interest! The oscillation of spatiality learning along the x- and the y-axis could be caused by the occasionally occurring inconsistent language label supervision. If you look at the dataset, you would find that for some objects, the collected labels are sometimes contradicted in descriptions about "left/right" and "front/behind" (but there are no such issues for "below/above", so learning along the z-axis is stable). Hope this helps!

Best, Heng