Training Settings for FreiHAND

[BalPdk]

Hi! Thanks for providing such amazing works! I am trying to reproduce your results on the FreiHAND dataset, but the best I obtained is slightly lower than your provided checkpoints. My results for lixel stage are as follows:

Evaluation 3D KP results:
auc=0.231, mean_kp3d_avg=6.03 cm
Evaluation 3D KP ALIGNED results:
auc=0.848, mean_kp3d_avg=0.77 cm

Evaluation 3D MESH results:
auc=0.230, mean_kp3d_avg=6.04 cm
Evaluation 3D MESH ALIGNED results:
auc=0.846, mean_kp3d_avg=0.78 cm

F-scores
F@5.0mm = 0.130     F_aligned@5.0mm = 0.668
F@15.0mm = 0.395    F_aligned@15.0mm = 0.971

I have modified the config to use FreiHAND (python trainset_3d: ['FreiHAND'] trainset_2d: [] testset: FreiHAND) and batch size to 32 since I am using 2 gpus. I was wondering:

1. Is this result I obtained normal based on your experience? Do you think of any places I should pay attention to?
2. If it is normal, could you please offer me some suggestions to squeeze a bit more to reach the one you obtained?

Many thanks in advance!

[mks0601]

Hi, how many epochs did you train your network?

[BalPdk]

Hi, I trained for 25 epochs with LR decreasing at the 17th and the 21st epoch.

[mks0601]

Sorry , I haven't faced the same error :( Maybe because of the mismatch of the batch and gpu numbers?

[BalPdk]

I see. Let me have another try with your settings haha... Thanks :)

[BalPdk]

@mks0601 One quick question :), in your experiments, did you use FreiHAND V1 or V2?

[mks0601]

Actually, I have no idea.. Do you know when v2 became available? The results of this paper is obtained in the early of 2020.