Question about dataset

[unrealMJ]

Excellent work~

i have some questions about collecting the dataset.

1. the Miradata has about 475k video clips, which is much larger than the 10k dataset in the paper，is there some filtration？
2. the miMixamo has motions and characters， When rendering, is it only necessary to use a simple HDR, or is it required to place the characters in a complex 3D scene?
3. Could you tell me the specific ratio of real data to synthetic data?

Thank you in advance for any reply🙏

[skygoo2000]

Thank you very much for your questions!

1. We evenly selected different categories from MiraData's various class labels. The small size of the training dataset is primarily due to our limited resources.
2. We used Blender to render Mixamo's character animations and randomly set up skyboxes and ground textures.
3. The ratio of real data to synthetic data is approximately 7:3.

[unrealMJ]

Thank you very much for your quick reply！

I have two additional questions：

1. In rendering, does the camera use a fixed pose, or does it use a trajectory?
2. The MIRA dataset is publicly available. Could you share your subset? We hope to reproduce your results.

Thank you once again for your help.

[skygoo2000]

We are currently organizing our dataset, which may take some time. In the meantime, you can use generate_TrackingVideo.sh to create your own dataset by following the instructions in the README.

As for synthetic data, we randomly initialize a camera position within a fixed range and ensure that the camera always points toward the character.

[unrealMJ]

Thank you for your generous reply!

I want to ask two more questions about the caption.

1. The miradata has several captions, which do you use in das? or recaption the video clip？
2. Which model is used to caption the synthetic video？

Thank you once again for your help.🙏🙏