Help with Point Control Results

[byran-wang]

Hi Omni team,

Thank you for your great work. I tested the model with point control, but the results don’t look good. Could you please help me determine whether this issue is due to a bug on my side?

The input image and points are shown below:

The point control result (run with python3 inference.py --control_type point) is shown here:

You can find the input raw data and output results at the following link:
Google Drive folder

Thanks again for your support!

[LongHZ140516]

Hi @byran-wang

I tried using your data, and the code I ran was the original inference.py from the repository, with no modifications other than the input paths (image: 010_input.png, ply: 010_input.ply). The result I obtained looks like the image I’ve attached. I suspect it might not be an issue with the model itself. You might want to double-check your input files or your runtime setup to see if there’s any problem there.

[hanryyan]

For the point cloud condition, it is better to ensure that the coordinate system of the rendering image aligns with that of the partial point cloud and ensure the point cloud is Y-up.

[byran-wang]

Hi @LongHZ140516 and @HongyuYann,

I noticed that the code was updated after my initial test on September 28.

I’ve now tested the latest version (commit ID: 7c4d718) and followed @HongyuYann’s suggestion to set the Y direction of the point cloud as upward. The results are shown below:

However, the generated car result is not well aligned with the input point cloud, and the generated cake result contains many holes.

You can find the input raw data and output results at the following link:
🔗 Google Drive – Test Data and Results