Open qingchenkanlu opened 2 months ago
@qingchenkanlu Have you solved this problem?
I replace dataset_pair to dataset, the train code can run. However, I cannot fully reproduce the results presented in the paper. 2024-07-08 15:07:37,689 - **** Start Logging **** 2024-07-08 15:13:15,914 - 3D IoU at 25: 84.2 2024-07-08 15:13:15,915 - 3D IoU at 50: 83.9 2024-07-08 15:13:15,915 - 3D IoU at 75: 75.6 2024-07-08 15:13:15,915 - 5 degree, 2cm: 52.6 2024-07-08 15:13:15,915 - 5 degree, 5cm: 63.5 2024-07-08 15:13:15,915 - 10 degree, 2cm: 68.5 2024-07-08 15:13:15,915 - 10 degree, 5cm: 86.3 2024-07-08 15:13:53,300 - 3D IoU at 25: 84.0 2024-07-08 15:13:53,300 - 3D IoU at 50: 80.8 2024-07-08 15:13:53,301 - 3D IoU at 75: 40.5
If someone manages to reproduce the results of the paper in the future, they can provide the pre-trained weights.
: 52.6 2024-07-08 15:13:15,915 - 5 degree, 5cm: 63.5 2024-07-08 15:13:15,915 - 10 degree, 2cm: 68.5 2024-07-08 15:13:15,915 - 10 degree, 5cm: 86.3
Thanks for your information! Did you run VI-Net before? If so, could you reproduce the training results of VI-Net? It seems the ts network in this work is the same as that in the VI-Net work.
No, I have not. VI-Net provides complete experimental results and training weights. I consider the results to be reliable. SecondPose is developed based on VI-Net, ts network is similar.
Well, I could not reproduce the training results of VI-Net, although their provided weights have close results to their paper.
No module named 'dataset_pair'. The dataset_pair.py has not been uploaded?