Incorrect FPS with synthesized IMU?

[Yefancy]

https://github.com/SPICExLAB/MobilePoser/blob/main/mobileposer/process.py#L15

TARGET_FPS = 30

according to your code, I noticed that you sampling the motion to 30hz

however, the acceleration generation is 60Hz
https://github.com/SPICExLAB/MobilePoser/blob/main/mobileposer/process.py#L23-L32

def _syn_acc(v, smooth_n=4):
    """Synthesize accelerations from vertex positions."""
    mid = smooth_n // 2
    acc = torch.stack([(v[i] + v[i + 2] - 2 * v[i + 1]) * 3600 for i in range(0, v.shape[0] - 2)])
    acc = torch.cat((torch.zeros_like(acc[:1]), acc, torch.zeros_like(acc[:1])))
    if mid != 0:
        acc[smooth_n:-smooth_n] = torch.stack(
            [(v[i] + v[i + smooth_n * 2] - 2 * v[i + smooth_n]) * 3600 / smooth_n ** 2
             for i in range(0, v.shape[0] - smooth_n * 2)])
    return acc

I double check the transpose codebase, they were using the 60hz both motion and acceleration by default.
https://github.com/Xinyu-Yi/TransPose/blob/main/preprocess.py

I'm not sure whether it will cause a bug?

[GaigeY]

Hi Ye, I am also working on validating the results of MobilePoser using my own motion capture dataset. I followed the AMASS processing script provided by the authors, similar to IMUPoser, to obtain the data for comparison.

However, I have encountered some similiar issues with the reconstruction results. I noticed that while you mentioned the results were not satisfactory, my findings have been quite similiar. I tested both 30Hz and 60Hz versions (modified with TransPose and IMUPoser), but both yielded high PA-MPJPE values exceeding 130 mm, whereas I observed lower values with TransPose (90+) and PIP (around 50+).

I am unsure if there might be an issue with my replication steps. Could you please share your replication results? I would greatly appreciate any insights you might have regarding your experience with MobilePoser.

Thank you for your time!

[sinchaoJa]

Hi Ye, I am also working on validating the results of MobilePoser using my own motion capture dataset. I followed the AMASS processing script provided by the authors, similar to IMUPoser, to obtain the data for comparison.

However, I have encountered some similiar issues with the reconstruction results. I noticed that while you mentioned the results were not satisfactory, my findings have been quite similiar. I tested both 30Hz and 60Hz versions (modified with TransPose and IMUPoser), but both yielded high PA-MPJPE values exceeding 130 mm, whereas I observed lower values with TransPose (90+) and PIP (around 50+).

I am unsure if there might be an issue with my replication steps. Could you please share your replication results? I would greatly appreciate any insights you might have regarding your experience with MobilePoser.

Thank you for your time!

Hello, I used the weights.pth file provided by the author for evaluation. The experimental results suggest that the weights.pth file was most likely trained under 30Hz conditions, rather than 60Hz. This is my hypothesis.

[sinchaoJa]

Hi Ye, I am also working on validating the results of MobilePoser using my own motion capture dataset. I followed the AMASS processing script provided by the authors, similar to IMUPoser, to obtain the data for comparison.

However, I have encountered some similiar issues with the reconstruction results. I noticed that while you mentioned the results were not satisfactory, my findings have been quite similiar. I tested both 30Hz and 60Hz versions (modified with TransPose and IMUPoser), but both yielded high PA-MPJPE values exceeding 130 mm, whereas I observed lower values with TransPose (90+) and PIP (around 50+).

I am unsure if there might be an issue with my replication steps. Could you please share your replication results? I would greatly appreciate any insights you might have regarding your experience with MobilePoser.

Thank you for your time!

I recently tested the dip_test.pt and imuposer_test.pt files generated using the process.py script provided by the author. With the weights.pth file and the ONLINE setup, I evaluated using the MPJVE metric. The results are as follows: imuposer: 9.44, dip: 12.55.
According to the author's paper, the fine-tuned results for imuposer should be 9.5 and for dip 12.2.
I did not experiment with the TotalCapture dataset, as I am unclear about the author's data processing approach for this dataset.
Would you be willing to share your reproduction process with me?

[taeyoungyeon]

updated code for reflecting fps for calculation. Try again and whether issue was solved. Thanks