Closed nitba closed 5 years ago
Hi @Ivamcoder, I believe I just ran eval3d.py from the provided code of Zimmermann & Brox and got this number. It may be entirely possible that there may be minor discrepancies due to version differences. What number do you get?
Hi @spurra ,
This is what I get, so far from 8.56
,
One more question is that in Table2
you have omited Handness
which represent the left or right hand,
and the scale
,
scale Augmentation for data and keeping the hand bounding box in a range
? if so would you please confirm me that I get correctly.and
if we uniform the size of hand bounding box mentioned in paper as input into network how can we say the output of network has correct scale to be scale invariance , I got confused :(
when you have used both view of STB dataset , it means that instead of 15K
images for training , you have 30K
images?
Hi @spurra ,
This is what I get, so far from
8.56
,
Is this using the ground truth hand cropped images?
One more question is that in
Table2
you have omitedHandness
which represent the left or right hand, and thescale
,* How you omit the scale , Does it mean that the output hand pose is just palm-relative but not normalized respect to a reference bone? so in this experiment did you use `scale Augmentation for data and keeping the hand bounding box in a range`? if so would you please confirm me that I get correctly.
and
* if we uniform the size of hand bounding box mentioned in paper as input into network how can we say the output of network has correct scale to be scale invariance , I got confused :( * when you have used both view of STB dataset , it means that instead of `15K` images for training , you have `30K` images?
Hi @spurra ,
Is this using the ground truth hand cropped images?
yes, in eval3d.py
, I have BinaryDbReaderSTB(mode='evaluation', shuffle=False, hand_crop=True, use_wrist_coord=False)
I'm not quite sure where the discrepancy comes from. If I find the time, I'll investigate this.
Thank you @spurra
Hi @Spurra, Inreagrd with table 2. in your paper, how you get the number of "8.68" from [38] work, in paper I did not found such a number and when I run his 'eval3d.py' provided script I got different number,
Would you please tell me the setup for his code you have used to get this number?
Thanks,