search

ethnhe / PVN3D

Code for "PVN3D: A Deep Point-wise 3D Keypoints Hough Voting Network for 6DoF Pose Estimation", CVPR 2020
MIT License
492 stars 105 forks source link

Occ-linemod #74

Open hillaric opened 3 years ago

hillaric commented 3 years ago

First, congratulation for your new work FFB6D. I see that you test the occ_linemod in both methods, and I want to test PVN3D on occ_linemod, but I don't have the preprocessed occ-linemod data. In detail, the downloaded dataset can not be used directly, how do you process the data?

ethnhe commented 3 years ago

We use preprocessing code from PVNet to transform the pose parameter (pvnet_code) and then use similar scripts as in our datasets/linemod/linemod_dataset.py to generate input to our model.

hillaric commented 3 years ago

thank you! I will have a try.

hillaric commented 3 years ago

The 'blender K' in common.py is the same K used in the 'occ_linemod'? I am a little confused.

ethnhe commented 3 years ago

No, both the linemod and occ_linemod use 'linemod' K, the 'blender' K is only used for the synthetic dataset.

hillaric commented 3 years ago

Would you mind provide the corresponding code, especially the gt.yml. I try, but I can't do it. image

hillaric commented 3 years ago

After using the code 'Pose_transformer' you released in 'base_utils.py' for OCCLUSION_LINEMOD, I got a different pose from PVNet. Is it right?

ethnhe commented 3 years ago

You can check if the pose is correct by transforming the object vertexes or the selected keypoint by the pose parameters and project them back to the RGB image for visualization, as what we have done in datasets/linemod/linemod_dataset.py code link.

aiai84 commented 3 years ago

hello, have you solve the problem, I want to test the OCC-Linemod dataset, could you show your change or addition on the original code? Thank you

William-Klay commented 2 years ago

For occ-lm, i found the model performance on eggbox is pretty bad. Can you please give me any insight about it?

d-serafly commented 2 years ago

For occ-lm, i found the model performance on eggbox is pretty bad. Can you please give me any insight about it?

May I know how you completed the test on occ? Especially in occ dataset processing.

weidu3 commented 1 year ago

For occ-lm, i found the model performance on eggbox is pretty bad. Can you please give me any insight about it?

Yes, I also found this problem, eggbox in the LM-O dataset has been suppressed by a red car, and I can't reach the add-s in the article with the pre-training weights provided by the author or training myself. So I'm curious, eggbox is testing on the LM-O dataset, is it using the original data of LM-O?