wzzheng / pointocc Goto Github PK

Create soft link from data/nuscenes to your_nuscenes_path

Thanks for your work! However, I modified the parameter configuration in config to reduce the resolution of voxel features as recorded in section 4.2 of the paper.
PointOcc/config/pointtpv_nusc_occ.py
Default: grid_size_occ = [512, 512, 40] Modified: grid_size_occ = [256, 256, 20]
The following error occurred during dataset construction:
IndexError: index 354 is out of bounds for axis 1 with size 256

So I need to re-generate the occupancy_nuscenes dataset, what is the benchmark of the dataset that we used in the paper? Could you provide the relevant conversion code? Thank you very much.

When I use the torch. onnx. export command to export the ONNX model, an error which the type of grid_ind is NoneType appears. I am not sure what its definition is and how the parameters are passed into the forward

File "~/miniconda3/envs/pointocc/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py", line 287, in _worker_loop
data = fetcher.fetch(index)
File "~/miniconda3/envs/pointocc/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 52, in fetch
return self.collate_fn(data)
File "~/project/PointOcc/dataloader/dataset_wrapper.py", line 303, in occ_custom_collate_fn
grid_ind_stack = np.stack([d[2] for d in data]).astype(np.float32)
File "<array_function internals>", line 200, in stack
File "~/miniconda3/envs/pointocc/lib/python3.8/site-packages/numpy/core/shape_base.py", line 464, in stack
raise ValueError('all input arrays must have the same shape')
ValueError: all input arrays must have the same shape

Thanks for your work! However, I find this part of code a little confusing

That's amazing!
I've succeeded up to eval_occ. However, I'm having trouble visualizing occ.
How can I do that? Please help.

Thank you for your work！

I'm examining the parameters max_volume_space = [50, 3.1415926, 3] and min_volume_space = [0, -3.1415926, -5], which define the maximum and minimum volume space for processing data. After converting these to arrays with np.asarray, I calculate the range of each dimension as crop_range = max_bound - min_bound, resulting in [50, 6.2831852 (2π), 8].

My questions are:

1. How were the ranges for the radius and height ([50, 2π, 8]) determined? Is there a specific rationale or dataset characteristic that guided these choices?

2. Regarding the line xyz_pol_grid = np.clip(xyz_pol, min_bound, max_bound - 1e-3), I'm concerned about the potential modification of the original point positions. By clipping with max_bound - 1e-3, does it not alter the spatial coordinates slightly, especially for points near the boundary? What's the intent behind this adjustment?

Thanks for your work！Could I have the occ pickle files for study？