search

hongsukchoi / TCMR_RELEASE

Official Pytorch implementation of "Beyond Static Features for Temporally Consistent 3D Human Pose and Shape from a Video", CVPR 2021
MIT License
287 stars 39 forks source link

Query about Evaluation Strategy #47

Open AmoghTiwari opened 7 months ago

AmoghTiwari commented 7 months ago

I noticed that in the evaluate.py file, in lines 119-123, for repr_table4_h36m_mpii3d_model case h36m_test_25fps_db.pt is used. But for the repr_table6_h36m_model case, h36m_test_front_25fps_tight_db.pt is used. Can you please clarify what is the difference between the 2 .pt files? And why two different files have been used ?

hongsukchoi commented 7 months ago

The differences are 1) the first pt includes all 4 camera images and 2) the second pt includes only frontal camera images and the bbox used for cropping is tight.

Table 4 is training and testing with our setting which is applied to all methods for fairness. The bbox tightness is similar across all datasets.

Table 6 is following the convention, mainly from SPIN. It uses a tight bbox for h36m images.

Best regards, Hongsuk Choi https://hongsukchoi.github.io

On Fri, Jan 26, 2024 at 12:30 PM Amogh Tiwari @.***> wrote:

I noticed that in the evaluate.py file, in lines 142-146, for repr_table4_h36m_mpii3d_model case h36m_test_25fps_db.pt is used. But for the repr_table6_h36m_model case, h36m_test_front_25fps_tight_db.pt is used. Can you please clarify what is the difference between the 2 .pt files? And why two different files have been used ?

— Reply to this email directly, view it on GitHub https://github.com/hongsukchoi/TCMR_RELEASE/issues/47, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHNGKW4BN2SQK7THANGU4ETYQPR3HAVCNFSM6AAAAABCMPLZMCVHI2DSMVQWIX3LMV43ASLTON2WKOZSGEYDENJZGAZTGNI . You are receiving this because you are subscribed to this thread.Message ID: @.***>

GloryyrolG commented 7 months ago

Hi @hongsukchoi @mks0601 , May I ask if u inspect or have insights on the flipped global orientation problem? I would have expected it not to happen cuz of learned temporal smoothness. Thx & best,

000013

hongsukchoi commented 7 months ago

Hi!

Thank you for your interest. Sometimes TCMR can fail of course, but this result looks interesting, because it shows a good pose but with a wrong global orientation.

My guess is

1) Just a failure due to an extreme posed person image. 2) The 3d mesh is accurate, but the projection is wrong. You may want to check the 3d mesh sequence directly.

Best regards, Hongsuk Choi https://hongsukchoi.github.io

On Wed, Feb 21, 2024 at 12:06 PM GloryyrolG @.***> wrote:

Hi @hongsukchoi https://github.com/hongsukchoi @mks0601 https://github.com/mks0601 , May I ask if u inspect or have insights on the flipped global orientation problem? I would have expected it not to happen cuz of learned temporal smoothness. Thx & best,

000013.png (view on web) https://github.com/hongsukchoi/TCMR_RELEASE/assets/33202229/157dd1be-4669-4b08-9d89-9b145066336f

— Reply to this email directly, view it on GitHub https://github.com/hongsukchoi/TCMR_RELEASE/issues/47#issuecomment-1957328657, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHNGKW7XOZ6DQA4TNJBTNI3YUYSQNAVCNFSM6AAAAABCMPLZMCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNJXGMZDQNRVG4 . You are receiving this because you were mentioned.Message ID: @.***>