fullcyxuc / b-seg Goto Github PK

Hi, thank you for your great effort. I would like to know whether your dataset contains accurate camera intrinsic and extrinsic parameters for each image. If so, it will be very suitable for MVS and NeRF applications. I have checked the paper and didn't find clear information about this. Could you give me some hints? Thank you very much.

Can you release the description of the datasets?
The datasets have several columns and can you give the description of each column especially the fine-grained building labels

Thanks for the wonderful work! I wonder how long does it take to train with single 3090 GPUs?

Hi, Thanks for sharing this great work!
Could you please estimate the code release date?
Thanks

I'm having some trouble trying to reproduce the algorithm. I wonder if it's my environment configuration that's having some issues.
The environment I use is basically according to the configuration requirements in git: python version 3.6.2, pytorch is 1.2.0, cuda version 10.0, graphics card is 3080Ti (12GB), but I still get cuda execution failed with error 500 when I call the torch.ops.spconv.get_indice_pairs_3d() function at the start of training

The function with the error is：torch.ops.spconv.get_indice_pairs_3d()

I initially suspected that it was the graphics card that there was not enough memory, tried training with the smallest Lihu dataset and scaling down the batch_size and max_npoint, and still didn't solve the problem.