epic-kitchens / epic-fields-code Goto Github PK

Hi team,

Great work! Congratulations!

Could you please explain the relation between the colmap commands in Issue#3 and the raw COLMAP format from EPIC-FIELDS website? More specifically, the two links on the website the output or the input to the 3 colmap commands? Thanks!

The dense registered frames in raw COLMAP format can be found at: here (133G).
The sparse frames including raw COLMAP database can be found at: here (91.6G).

Originally posted by @zhifanzhu in #3 (comment)

The filtering step selects a subset of image frames for each video, over which camera poses are computed. How were the camera poses that are released under EPIC-FIELDS computed? They are far denser than would be allowed by the image selection filter.

Hi! While your work and dataset are truly inspiring, I'm wondering how can we obtain the very dense point cloud (illustrated in the header image, the paper's figure1 or the video you posted) from either sparse reconstruction or dense registration in COLMAP. Either the sparse/dense raw data you made available is far more sparse than in the figure.

I will be more than appreciated if you could illustrate the process since a vivid and expressive point cloud background can be truly helpful in all ways.

Thank you!

Dear authors, thank you for your amazing work!

Are there any plans to release the code used to perform the "filtering" step described in your work, such that out-of-dataset videos can also be processed into point clouds using your method? If so, when is the code expected to be released?

Thank you in advance!

Hi! I want to reproduce the single-image 3D hands (FrankMocap) integration in EPIC-Fields visualization, as shown in the bottom row of Figure 2 of the paper "EPIC Fields: Marrying 3D Geometry and Video Understanding".

Do you guys plan to release the code for producing this visualization?

Also, it would be helpful if you can share some details of how you achieved the integration. FrankMocap produces 3D hand pose in either the SMPL coordinate or the image coordinate (where depth is ambiguous). What method did you guys use to convert FrankMocap's prediction to world coordinate? I assume that since the dataset contains camera poses in world coordinate, we can find FrankMocap's predictions in the camera coordinate, and then apply the camera2world transform. But how did you guys convert FrankMocap's predictions in SMPL coordinate or image coordinate to camera coordinate?

Many thanks in advance!

I'm interested in using EPIC-Fields and wanted to examine the paths of the camera-wearer in the 3D space for a set of videos after seeing the paper and visualizations. Is there any code to replicate those visualizations directly?

Hi, thanks for sharing this great work and data. I'm trying to run dynamic rendering on the given data, but I'm wondering whether the provided raw colormap data has distortion or not. How can i get the undistored images for rendering rays, thanks!