search

zju3dv / SMAP

[ECCV 2020] SMAP: Single-Shot Multi-Person Absolute 3D Pose Estimation
Apache License 2.0
241 stars 37 forks source link

training loss too large #13

Closed leonzgtee closed 3 years ago

leonzgtee commented 3 years ago

when i start training with this project, i find that the loss is too large to converge. after 70k iteration, loss_2d loss about 384,loss_bone about 200,loss_root about 2.is there anything wrong with these loss. thanks.

raypine commented 3 years ago

After 40k iteration, loss_2d is about 110. loss_bone is about 1. loss_root is about 0.1. The unit is cm in our setting. Is it correct ?

minhpvo commented 3 years ago

@leonzgtee Is it possible for you to share the json conversion script from MuCo to SMAP format? Thanks!

leonzgtee commented 3 years ago

@raypine the unit previous uesed was mm, after changed to cm, it seems very close to the value. now the loss_2d: 157.297, loss_bone: 1.538, loss_root: 0.078 after 100k iteration's training.

leonzgtee commented 3 years ago

@minhpvo i modified it from lib/preprocess/create_annot.py, just as follows: create_MuCo_annot.txt