davrempe / caspr Goto Github PK

Thanks for sharing good work.
I have some questions related to your work.

Q1. Have you tried experiments on NOCS-REAL 275 dataset [1] using propose the T-NOCS??
The representation of NOCS comes from the [1] paper and the dataset consists of video sequences dataset.
If I were you, I would have considered experiments with an existing real world video dataset (ex, NOCS-REAL 275).
If you did, I'm curious what difficulties you encountered. If not, I wonder why you didn't consider the experiment.

Q2. Can you explain the difference between GT-NOCS and GT-TNOCS??
May I understand that GT-TNOCS generated by the union of K partial of GT-NOCS??

Q3. In the limitations and future work part, It mentioned that "dense supervision of T-NOCS labels is may not available for real data".
As I understood, the NOCS-REAL training dataset [1] already has T-NOCS labels.
Can you explain clearly why you mentioned that may not be available for real data??
Have I missed something??

[1] Wang, He, et al. "Normalized object coordinate space for category-level 6d object pose and size estimation." CVPR. 2019.

• NOCS-REAL 275: evaluation set
• NOCS-REAl training dataset: training set

Hi @davrempe,

Could you provide the code for preparing data (the 3 objects and also the warping car)?

I want to create more data to test.

Thanks

Hi,

Thank you for your great work.
I have tried to reproduce the result by training from scratch, but I cannot obtain the result reported in the paper. (Median Chamfer around 3 vs 0.5 in the paper.)
I used batch size 36 on 3 A6000 GPU. My pytorch version is 1.7.1 with CUDA11. The rest hyper-parameters are the same.

Do you have any hint to reproduce the results?
Thanks.

Hi Davis, I created a docker image so that I can build the same environment as you mentioned here in the GitHub.
When I was about to run this command,
python viz.py --data-cfg ../data/configs/demo.cfg --weights ../pretrained_weights/caspr_weights_cars.pth --seq-len 10 --num-pts 2048 --viz-tnocs --viz-observed --viz-interpolated

The output in the terminal is