Baseline Performance Question #32
Description
Hello, It's me again!
I am currently trying to use the pretrained ZeroFlow model provided by the authors with the FastFlow3D model in OpenSceneFlow, using
/home/jang/eccv26/OpenSceneFlow/tools/zero2ours.py.
However, zero2ours.py does not work out of the box. After investigating the issue, I found the following discrepancy:
- In the official ZeroFlow code, the FastFlow3D implementation does not include offset_encoder in self.head.
- In contrast, the FastFlow3D implementation in your GitHub repository does include offset_encoder inside self.head.
Because of this structural mismatch, the converted checkpoint cannot be directly loaded.
To proceed, I modified your FastFlow3D code to match the official ZeroFlow implementation (i.e., removing offset_encoder from self.head).
Then I ran inference using the checkpoint "nsfp_distilatation_speed_scaled_updated_run2.ckpt."
However, the resulting performance is significantly lower than expected, as shown in the attached results.
I would like to ask for your advice!
For reference, I am evaluating on an older version of the dataset as GT, not the HIMO dataset.
Any guidance or suggestions would be greatly appreciated.
Thank you very much for your time and help!