Generate Training data

[heiheishuang]

Hello, thank you very much for your code. I am very interested in your work and I have tried using evaluate.py to generate nuScenes training data. But, I have noticed that its speed is a bit slow (only one GPU is in use). May I ask if there is any solution?

I tried to modify a single GPU to multiple GPUs, but failed: all GPUs are processing the same scene. May I know how to modify this code? Looking forward to your reply!

[heiheishuang]

Hello, I often get IndexError: list index out of range errors when generating NuScenes training data. Usually, they look like this
Is my situation normal?

It looks like it's because of scene_depth_results is empty. And the code about updating 'scene_depth_results':

How should I modify this code?

[AronDiSc]

Hi @heiheishuang , thank you for your interest in our work and sorry for the delayed response.

About the speed of generating the data: I've just used a single GPU to generate the data and it took me a couple of days. I guess if you have multiple GPUs available you could try to modify the code such that you can specify which gpu should be used. Then you could divide all nuScenes scenes into n split-files and n config-files where each file would look as follows:

nuScenes_split_gpu0.txt:

scene-0000
scene-0000
...

dataset_generation_nuscenes_gpu0.yaml:

...
datasets:
  validation:
     split: [nuScenes_split_gpu0.txt]
...
...

...

you could then run the script simultaneously multiple times for each available gpu (assuming you also have sufficient memory):

python evaluate.py --config <path_to_config>/dataset_generation_nuscenes_gpu0.yaml

About the error you've encountered: Sorry for the trouble this caused. For the specific scenes where this error occurs, the algorithm fails to do any reasonable scene reconstruction (e.g. because the scene is static) and thus no estimations are accumulated. This in itself is not an issue because we do not want to train on such sequences anyway. It can be fixed by simply changing L174 in eval_dataset.py to ...

if 'depth' in self.metrics and len(scene_depth_results) >= 1:

... i.e. we skip evaluation for scenes where we do not have any predictions. I've changed the code accordingly.

Note about NuScenes training: If you generate the data for NuScenes from scratch I suggest you remove L138 in NuScenesDataset.py. We actually had a bug in our training where the intrinsics we used for NuScenes were slightly off. Thus this line is there to get the same results as in the paper. You might be able to achieve better results by using the correct NuScenes intrinsics.

[VergettLee]

Hello, I have modified this code: if 'depth' in self.metrics and len(scene_depth_results) >= 1:
But a new error has appeared.

"I'm training on the DDAD dataset, but encountered an error similar to the one shown in the image. Initially, I suspected that the scene 000199 in my dataset was downloaded incorrectly. So, as shown in the image, I used scene 000198 as the last one, but still encountered the error 'IndexError: index 3906 is out of bounds for dimension 0 with size 3900'. What should I do to modify the code?"