Jianyuan Wang comments

Results 238 comments of


                                            Jianyuan Wang

Duplicate point cloud per camera

Hey this is a know problem for the released ckpt, that it sometimes tends to think there is a fake motion, especially for people or statue (even though the object...

Clarification needed: Are frame and global attention each complete transformer blocks (attention + MLP)?

Global and frame attention both use the full attention block. They do not share.

Fine-tune training

I would suggest using 8 GPUs for fine-tuning

Can not reproduce the online interactive result in local demo.py

Hey I think they should not have a huge difference. But just in case, you can clone our hugging face demo to local and run it. The code of the...

Data Prepare

We labelled it by ourselves. Cut3r has provided their own processed ones, as can be found here, if you want to use (I did not check their quality before though)...

extri_opencv should to camera_to_world or camera_from_world

Hey extri_opencv is camera from world

Which resolution downloaded from the dl3Dv dataset is used for training?

It should be 960p

Cannot download the model from Hugging Face

I duplicated one by myself. Can you try this? https://huggingface.co/spaces/JianyuanWang/vggt

Cannot download the model from Hugging Face

now https://huggingface.co/spaces/facebook/vggt has been fixed

Depth Accurary vs. World Point Accuracy, which is better?

Generally the point cloud achieved by unprojecting depth maps is better than the point map