Using data from '3d-pose-baseline'

[shrubb]

Hi again, In an earlier issue, you wrote that it should be possible to reproduce your results using 2D poses for Human3.6M provided in the 3d-pose-baseline repository. However, I was only able to achieve 53 mm and 71 mm error using ground truth and stacked hourglass 2D predictions respectively (expected 39.2 mm and 51.9 mm): This is with default settings (100 epochs, initial learning rate 1e-5 etc.), I only set the dropout rate to 0.5 and changed the data directory. Am I missing something or this is expected?

[rayat137]

Hi @shrubb What flags did you use? Did you use the flag --camera_frame? Sorry that I have not finished the documentation yet. I will finish working on it. Use the following command python temporal_3d.py --use_sh --camera_frame --dropout 0.5

[shrubb]

No, I didn't. I'll let it train with this flag overnight and check for the improvement tomorrow morning.

Thanks, your effort is awesome!

[shrubb]

Woohoo, that worked, thank you again! :tada:

[rayat137]

Thanks @shrubb. Glad it has worked so well. In fact the results look better than what I reported in the paper with the 2D poses provided in the 3d-pose-baseline repository.