BIWI Dataset about 6drepnet HOT 7 CLOSED

Hi,

we used 256 of the initial size to be able to augment the images by random cropping to its final size of 224. I suggest to make sure that the data preparation is done and no axis are accidentally switched. https://github.com/thohemp/6DRepNet/blob/master/sixdrepnet/datasets.py#L177

from 6drepnet.

Thanks for the response,

I have identified axes using Your demo code. Im processing the BIWI dataset with this code too. I'm loading BIWI_70_30 model. I have noticed, that there are many persons on some pictures, thus i pick the area with the biggest face for each image.
The only difference from https://github.com/shamangary/FSA-Net/blob/4361d0e48103bb215d15734220c9d17e6812bb48/data/TYY_create_db_biwi.py#L40 is that in the demo code Retina Face detector is used instead of MTCNN. Results I obtain are:

X Y Z corresponds to Pitch Yaw Roll respectively

from 6drepnet.

What evaluation script do you use? I just tested this repo again and it gives me reported results:

python test.py --dataset BIWI --filename-list /path/to/BIWI_70_30_test.npz --snapshot 6DRepNet_70_30_BIWI.pth

Output:

Yaw: 2.6911, Pitch: 2.9193, Roll: 2.3557, MAE: 2.6553

from 6drepnet.

Im using BIWI from HuggingFace. Im not using this preprocessing step from FSA-Net.
Steps Im doing:
For each image in the BIWI:

1. from ground truth rotation matrix compute X Y Z angles.
2. Crop face with Retina Face detector
3. Add 0.4*size offset
4. Make predictions with 6DRepNet_70_30_BIWI.pth
5. Compute absolute error for each angle.
   At the end, I compute MAE.

Im doing this by modyfing demo.py to not to work with webcam, but with the BIWI.
Does my results mean I should retrain 6DRepNet with the BIWI dataset preprocessed in my way to obtain results simiar to Yours?

from 6drepnet.

I haven't used the BIWI data from HuggingFace. Hence, I can't tell where the problem lies. But the huggingface data won't have the same split as mine, so, by using my model, you are probably testing on part of the data the model is already trained on. So, either stick with my way on data preparation or completely retrain on a new model on your data.

from 6drepnet.

I have followed Your steps and i got results as below:
Yaw: 1.6349, Pitch: 1.8322, Roll: 1.7049, MAE: 1.7240.
However, I have found splitting script to be non deterministic, thus probably some of test data have been seen durning training.
Do You mind sending me Your test part so I can compare the results with other algorithms?
[email protected]

from 6drepnet.

I won't share the dataset to not come in conflict with the corresponding license. Beyond that, my test data will give you different results, as your model is trained on different data. The script will provide you a dataset split, where training and test images are not the same. So it allows you to train the model yourself and generate similar results to mine.

from 6drepnet.