search

lschmidtke / shape_templates

code from cvpr 2021 paper 'Unsupervised Human Pose Estimation through Transforming Shape Templates'
https://infantmotion.github.io/
MIT License
32 stars 8 forks source link

Number of training images #10

Open unofgithub opened 1 year ago

unofgithub commented 1 year ago

Hello,

I just have a quick clarification about the number of images you used in training. You mentioned using about 600K images for training. Could you please let me know if that means it was trained with 300K pairs (frame 1, frame 2) or that 600K pairs were used for training? Thank you so much!

lschmidtke commented 1 year ago

Hi, it's 600k training images in total, but with random frame pairs.

unofgithub commented 1 year ago

Hello. Thank you. Could you please let me know how many random frame pairs were sampled?

lschmidtke commented 1 year ago

Okay, so I counted the number of possible pairs in each bounding box bucket for each subject and each activity. Number of possible pairs would be (num_frames choose 2). I end up with 800 million possible pairings. However, since the model is trained for a limited amount (around 30 epochs), there are roughly 30 epochs * 600k pairs seen during training. Keep in mind though that since it's video data, frame pairs that are close to each other in terms of timestamp have very little pixel-wise difference.

Hope that helps!