Unexpected EOF in archive error

[fattypiggy]

The tar file 002.tar -> ID40002.tar appears to be corrupted or incomplete. When extracting, it fails with an "Unexpected EOF in archive" error and is missing the timecodes/ directory that exists in all other sequences.

Expected Structure
Based on Subject 001 and the other sequences in Subject 002 (ID20002, ID30002), each sequence should have:

ID40002/
├── cameras/    (with 40 camera directories)
└── timecodes/  (with timecode .txt files)

Actual Structure：

After extraction, ID40002 only contains:

ID40002/
└── cameras/    (only 30 cameras, missing Cam07, Cam10, Cam13, Cam14, Cam19, Cam21, Cam26, Cam32, Cam35, Cam38)

Missing: timecodes/ directory

[prraoo]

@fattypiggy thank you for pointing that out, I’ve encountered the same issue. It seems that ID40002.tar is indeed incomplete, and a few other archives might have similar problems. I’m currently checking them.

In the meantime, please feel free to report any other broken sequences you find. I’ll work on fixing them as soon as possible.

[fattypiggy]

Yes, I've noticed a pattern: files with a size of around 60GB generally have no errors, meaning each unique ID (pose) is around 20GB. Files smaller than 60GB generally have errors. For files larger than 60GB, I found 4 unique IDs after decompression. Does this mean there are four poses? But doesn't the readme say three poses per person? I hope this information helps.

[fattypiggy]

Since you mentioned:
Step 1: Frame Extraction → High-quality EXR images (extraction/)
Step 2: Color Calibration → Color-corrected AVIF images (color-calibration/)
Step 3: Flow Alignment → Temporally aligned sequences (TODO)
Step 4: Camera Calibration and FLAME Tracking → Camera parameters and FLAME head model parameters (TODO)
Step 5: Relighting → Combining OLAT images to create relight images (TODO)

Do you think the results from the first two steps can be used in relighting? like only use AVIF images for training? I think this one without the camera location information or the environment maps. Would you consider provide these kind of information later?

[prraoo]

@fattypiggy Yes, your observations are correct.

Regarding the archives:
The ones around ~55 GB are indeed problematic. I’m currently working on fixing them and generating the correct versions.

Regarding Preprocessing:
For now, the extracted AVIF files are color-calibrated OLATs, which can already be used for face reflectance modeling. However, I agree that they are still limited by the available camera parameters and light directions.

I’ll do my best to release the remaining preprocessing components (excluding FLAME tracking) by the end of this week. In the meantime, here’s a quick heads-up on what’s coming: I use [RAFT](https://github.com/princeton-vl/RAFT) on fully lit frames for optical flow–based alignment and Metashape to obtain camera parameters. You could already start experimenting with those.

[fattypiggy]

I'm looking forward to see these new scripts and new files.
BTW: 068.tar missed.

[fattypiggy]

Hi @prraoo, Could you also provide the environment map and the camera location information?

[prraoo]

@fattypiggy yes, we do not have permission to release subject 068 (at least as of now). We will be adding new subjects.

Regarding camera calibration, I am currently working on it; unfortunately, it is the most involved process in our entire preprocessing pipeline.

[fattypiggy]

Thank you my friend, your work is absolutely amazing.
After reading the paper, I'm just curious about the way how to encode the view direction and light direction. Does the next step, step 4, camera calibration will include the environment map(or light direction, whatever)? Just take your time, good meals need some time.

[prraoo]

Haha, no worries ;)

The next step is to share the 3D light directions that are aligned with our light stage and then scripts for getting 3D meshes of the faces.

[fattypiggy]

I have some questions regarding this image:

1. What's the dimension of the env map at the right bottom corner? And what's the format of these?
2. How do you linearly combine two images together? As far as I know, only the HDR space is linearly addable, not for LDR space, like png or jpg files, what's the format are you using for training?
3. For the second row, I assume these are the complex environment, so the light can be combined by all the other "base" images?

I'm looking forward for your reply. THX.

[fattypiggy]

Hi @prraoo ,
I have a question regarding the radiometry of the raw data to ensure I am performing the relighting correctly:

I have located the cameras.calib file, but it seems to primarily contain camera intrinsics/extrinsics and color correction curves. I could not find specific radiance information for the light sources.

Question: Should we assume that all 331 light sources have a uniform white intensity (e.g., RGB 1.0, 1.0, 1.0)? Or is there a separate calibration file (like a lights.json or intensity.txt) that records the specific calibrated radiance for each light direction?

[prraoo]

@fattypiggy I have added sample relighting scripts; this should clarify your concerns. Please reach out to me if they are still resolved.