Impact by video resolution?

[tpm518]

Hello,

Thank you for this amazing work. I was able to get this running, and took a couple videos with the webcam on my macbook. The results are fantastic. When repeating the experiment with videos taken on my iPhone front-facing camera (selfie), the results don't work well. I'm assuming it is due to differing aspect ratios between the 2 videos. Any ideas on how to handle this?

Thanks!

[SoroushMehraban]

@tpm518 I also had the same issue with videos that height was more than width. I didn't have time to investigate but my instinct says it's probably due to the fact that it always sees videos from H3.6M dataset that is having square resolution while your IPhone videos are having higher height than width and here in demo when it normalizes the input it reaches a value for input that is unseen for model throughout training. I recommend adding padding to the input_2d prior to normalization as a initial step of debugging.