thohemp / 6drepnet360 Goto Github PK

Hi,

Could you put your full code? please.

I tried to implement your code, but I could not.

Thank you in advance!

Hi,

Could you make your fine-tuned model available, as you did in 6DRepNet? please.

I would like to test my own datasets on your own fine-tuned model for the comparison for narrow and full-range datasets.

Thank you in advance!

Thank you for your work!
I was trying to test the code on my own images. But it seems that I need .mat file for each image. I assume they are annoations of head. How can I create the .mat file? Could you please help with that? Or, is there any other way to use the code without .mat files.

Hello, I have a question about the method of calculating mae. The below is coded in your code.
"""
+= torch.sum(torch.min(torch.stack((torch.abs(p_gt_deg - p_pred_deg), torch.abs(p_pred_deg + 360 - p_gt_deg), torch.abs(
p_pred_deg - 360 - p_gt_deg), torch.abs(p_pred_deg + 180 - p_gt_deg), torch.abs(p_pred_deg - 180 - p_gt_deg))), 0)[0])
yaw_error += torch.sum(torch.min(torch.stack((torch.abs(y_gt_deg - y_pred_deg), torch.abs(y_pred_deg + 360 - y_gt_deg), torch.abs(
y_pred_deg - 360 - y_gt_deg), torch.abs(y_pred_deg + 180 - y_gt_deg), torch.abs(y_pred_deg - 180 - y_gt_deg))), 0)[0])
roll_error += torch.sum(torch.min(torch.stack((torch.abs(r_gt_deg - r_pred_deg), torch.abs(r_pred_deg + 360 - r_gt_deg), torch.abs(
r_pred_deg - 360 - r_gt_deg), torch.abs(r_pred_deg + 180 - r_gt_deg), torch.abs(r_pred_deg - 180 - r_gt_deg))), 0)[0])
"""
I don't understand why to use " torch.abs(p_pred_deg + 180 - p_gt_deg), torch.abs(p_pred_deg - 180 - p_gt_deg)".
For example, if p_pred_deg=10 and p_gt_deg=170, the error should be 160. But if I calculate it use your method, the error is 20.

Greetings, I was trying to run test.py, but could not download the model:

❯ python test.py
Loading data.
/home/mscherbina/Documents/github_repos/6DRepNet/dataset/AFLW2000/files.txt
Downloading: "https://cloud.ovgu.de/s/jFd5JAacLdAtZ4N/download/6DRepNet360_300W_LP_AFLW2000.pth" to /home/mscherbina/.cache/torch/hub/checkpoints/6DRepNet360_300W_LP_AFLW2000.pth
Traceback (most recent call last):
  File "/home/mscherbina/Documents/github_repos/6DRepNet360/sixdrepnet360/test.py", line 168, in <module>
    saved_state_dict = load_state_dict_from_url(
  File "/home/mscherbina/anaconda3/lib/python3.9/site-packages/torch/hub.py", line 727, in load_state_dict_from_url
    download_url_to_file(url, cached_file, hash_prefix, progress=progress)
  File "/home/mscherbina/anaconda3/lib/python3.9/site-packages/torch/hub.py", line 593, in download_url_to_file
    u = urlopen(req)
  File "/home/mscherbina/anaconda3/lib/python3.9/urllib/request.py", line 214, in urlopen
    return opener.open(url, data, timeout)
  File "/home/mscherbina/anaconda3/lib/python3.9/urllib/request.py", line 523, in open
    response = meth(req, response)
  File "/home/mscherbina/anaconda3/lib/python3.9/urllib/request.py", line 632, in http_response
    response = self.parent.error(
  File "/home/mscherbina/anaconda3/lib/python3.9/urllib/request.py", line 561, in error
    return self._call_chain(*args)
  File "/home/mscherbina/anaconda3/lib/python3.9/urllib/request.py", line 494, in _call_chain
    result = func(*args)
  File "/home/mscherbina/anaconda3/lib/python3.9/urllib/request.py", line 641, in http_error_default
    raise HTTPError(req.full_url, code, msg, hdrs, fp)
urllib.error.HTTPError: HTTP Error 404: Not Found

Thank you for sharing your great code. 😺
I have enjoyed it very much since the 6DRepNet generation. Thank you.

What is the license for this model? I'd like to cite it to the repository I'm working on if possible, but I want to post the license correctly.

https://github.com/PINTO0309/PINTO_model_zoo/tree/main/423_6DRepNet360

• My test results - Gold-YOLO (onnx) + 6DRepNet360 (50 batch / onnx)

  output_m50.mp4

Hello, The link 6DRepNet360_Full-Rotation_300W_LP+Panoptic.pth of seems to be invalid and cannot be downloaded.

Hello,
Good works, How to get headpose datasets (pitch\yaw\roll) of Panoptic ?

Hi @thohemp! Do you intend to make more updates on the implementation of the 6DRepNet360? I am currently searching for a model to infer the human head orientation on videos for a research project and the updates you have made on the previous 6DRepNet model is what I have been looking for. Thanks in advance.

I implemented your code using this CMU Panoptic. I got 2.66 as you obtained. The issue is in your test.py code. You should correct it and get MAE around 6.68.
I think if you will report this result. It will affect your first work (6d rotation representation for unconstrained head pose estimation).

I hope that I am wrong and you are correct.

Try to change the batch size =1 and see your results.

@thohemp

I am using the test script to run on my own video. However, the result is not very good as seen below.

could it be that the model available to download is not trained on CMU dataset?

Dear thohemp,

I used your programming demo to estimate head orientation. While I was able to measure up to 180° of the front of the face, I could not achieve 360° measurements. I would like to understand why this is the case.

I based my implementation on the sixdrepnet version of 6DRepNet (one version before 360°) and incorporated elements from 6DRepNet360.

Here is what I did specifically:

1. Added the SixDRepNet360 class (lines 25 to 84 of sixdrepnet360's test.py) to sixdrepnet's model.py.
2. Imported the SixDRepNet360 class into sixdrepnet's demo.py.
3. Specified the path to 6DRepNet360_Full-Rotation_300W_LP+Panoptic.pth in the default argument on line 41 (this file is the 6DRepNet360 Full Rotation model downloaded from Github's 6DRepNet360 Models).
4. Changed the model assignment on line 65 from SixDRepNet to SixDRepNet360, using the argument (torchvision.models.resnet.Bottleneck, [3, 4, 6, 3], 6).

My virtual environment includes:
・Python >= 3.9.19 (not specified in requirements.txt)
・Matplotlib >= 3.3.4
・Numpy >= 1.26.4 (original: 1.19.5)
・OpenCV-Python >= 4.10.0.82 (original: 4.5.5)
・Pandas >= 1.1.5
・Pillow >= 8.4.0
・Scipy >= 1.10.1 (original: 1.5.4)
・Torch >= 1.10.1
・Torchvision >= 0.11.2
For those dependencies with different versions, I tried to match the specified versions, but some versions were unavailable.

Other environment details:
PC Processor: AMD Ryzen 9 5900HX with Radeon Graphics 3.30 GHz
OS: Windows 10
GPU: NVIDIA GeForce RTX 3080 Laptop
Webcam: Brio 100
I would appreciate any guidance or suggestions on why the 360° measurement is not working and how I might resolve this issue.

Thank you for your assistance.

Best regards,
AonTitan

Greetings, I was wondering when the code will become available, I am particularly interested in conversion of CMU Panoptic dataset into Euler Rotation Degrees for heads.
You had mentioned in your paper that this was previously done by WHENet, but the original repo is some esoteric keras code. So I was wondering whether we can get pytorch implementation and conversion script.

Also, there is an oversight in your evaluation script, where you subtract and add 180°:
https://github.com/thohemp/6DRepNet360/blob/master/sixdrepnet360/test.py#L187

You probably copy pasted your old code. This does not affect current results (tested with AFLW), but may be a hidden footgun in the future, I suggest you fix that