Thanks for your great work!
I reproduced the results and noticed that the Gaussians for each pair of keyframes seem to be predicted independently, as shown in model/model_utils.py#L635. I also noticed that splat_inference.py#L241 directly skips previous/earlier key frame, which I assume is to avoid rendering the same frame repeatedly.
I understand that this design may make training easier. However, I am wondering how the Gaussian tracking results are obtained(as it seems that there is no gaussian maintained across all key frames). Are they computed by simply treating Gaussians at the same index/location in the Gaussian tensor across frames as the same Gaussian?
Thanks in advance!
Thanks for your great work!
I reproduced the results and noticed that the Gaussians for each pair of keyframes seem to be predicted independently, as shown in model/model_utils.py#L635. I also noticed that splat_inference.py#L241 directly skips previous/earlier key frame, which I assume is to avoid rendering the same frame repeatedly.
I understand that this design may make training easier. However, I am wondering how the Gaussian tracking results are obtained(as it seems that there is no gaussian maintained across all key frames). Are they computed by simply treating Gaussians at the same index/location in the Gaussian tensor across frames as the same Gaussian?
Thanks in advance!