How do I find the synced frames in the dataset

[bobluoluo]

Are datasets synced frame by frames? If not, How do I align the synced frames across the videos? I extracted the first frames from a set (coffee_martini). They don't look synced.

[zhaoyang-lv]

Thanks for flagging this. I am looking at this issue which likely is caused by extraction script. Will reply back later this week on this sequence.

[bobluoluo]

ok, Thanks. I was using opencv videoio. Let me try bare ffmpeg interface too.

[qingxu-thu]

Have the problem fix? By the way, the poses have only 18 views but the carmeras have 21, I wonder if some carmeras have not been included?

[zhaoyang-lv]

By the way, the poses have only 18 views but the carmeras have 21, I wonder if some carmeras have not been included?

We have discussed in the paper. We excluded all the cameras that not synced properly after final check in the timestamps.

[qingxu-thu]

Thanks for your reply, Have the details about which specific cameras are included and their corresponding poses been provided in your dataset and training? I wonder which specific videos and corresponding poses are used in your training. And I see that "camera poses correspond to their sequential sorted order in the folder". But in this coffee_martini scene, I can't identify the poses and cameras relationship since only 18 given. Can I use ffmpeg to exclude the unsyn ones and next the poses are sorted by the number of reamining videos?

[zhaoyang-lv]

@qingxu-thu does the readme answer the question? https://github.com/facebookresearch/Neural_3D_Video#neural-3d-video-synthesis-dataset

All camera should be synced in the release folder. Before the issue being sorted out specifically for the coffee_martini scene, please simply exclude this camera and its pose in the sorted index order.

[qingxu-thu]

Ok, I got it. Thank you very much. I know where I've made a mistake.

[snarb]

@zhaoyang-lv is this issue fixed? By the way what scene scale factor did you use? Can't find in paper