search

balakg / posewarp-cvpr2018

MIT License
194 stars 47 forks source link

Questions about the pipeline #2

Closed dulucas closed 4 years ago

dulucas commented 6 years ago

Hi Guha,

 Glad that you've shared your work, and I've tested on the Human 3.6M dataset, the results are magnificent!
 And here are some questions:
 1. For the loss function, why you choose the VGG instead of Resnet? Is there any benefits? (As I know, in the area of single image depth prediction, the depth images generated by VGG are more sharper of Resnet's)
 2. Have you tried on any difficult poses? the poses from extreme sports for exemple. 
  Thanks!
balakg commented 6 years ago

Great to hear!

  1. Resnet may also work fine. I chose VGG because it was used in some past work for texture synthesis (https://arxiv.org/pdf/1505.07376.pdf) and view synthesis (https://arxiv.org/pdf/1703.02921.pdf). The different blocks of VGG have also shown to correspond to progressively more complex image features, so you can choose blocks depending on what to emphasize. But presumably, resnet trained for image classification would also work.

  2. The poses we tried were limited to fitness, golf and tennis videos, so no, we haven't tried it on more extreme sports yet. Though I think some of the poses in our videos can be quite complex.

dulucas commented 6 years ago

Thanks a lot ! What if there are some key points are missing?

babyjia commented 6 years ago

Will,I would like to know that where is the training dataset?

dulucas commented 6 years ago

http://vision.imar.ro/human3.6m/description.php Here is the H3.6M dataset

babyjia commented 6 years ago

Ｉ am sorry to bother you that I coun't find the test code

balakg commented 6 years ago

@lucasdu007 You mean if there are joints missing?

dulucas commented 6 years ago

@balakg yes, if there are joints missing, what should I do?

balakg commented 5 years ago

If there are joints missing, those heat maps remain blank.