Daniel-xsy / RoboBEV

RoboBEV: Towards Robust Bird's Eye View Perception under Common Corruption and Domain Shift
https://daniel-xsy.github.io/robobev/
315 stars 32 forks source link

Generating nuscenes-c dataset #13

Open leesunmyung opened 1 month ago

leesunmyung commented 1 month ago

Hi, thanks for your project.

I downloaded nuScenes-C dataset, but it contains only 6,019 images for each camera.

I want to generate fog, snow dataset for all nuScenes training data using your code.

Could you tell me which code I have to use?

Daniel-xsy commented 1 month ago

Hi, thanks for your interest,

The nuScenes-C dataset is generated using the validation set of the nuScenes dataset, which only contains 6019 images for each camera. You can refer to this script for the corruption generation.

However, if you want to generate corruption for training data, it's recommended to generate corruption online instead of offline since it takes a lot of time and storage. You can try to integrate the script into the training data preprocessing pipeline.

leesunmyung commented 1 month ago

Thanks for your reply.

Due to my project, I cannot help but synthesize fog and snow image offline using the training data.

In CREATE.MD, you said to specify the data path carefully in config and specify the corruptions parameters, but there is no content in corruptions/project/config/config.py.

Could you tell me how to specify the data path and corruptions parameters? Thank you.

image

Daniel-xsy commented 1 month ago

Sorry for the confusion, you can refer to corruptions/project/config/nuscenes_c.py

Please also notice that this script is used to generate corruption for nuScenes dataset. For your need to generate corruption on customized dataset, I'm sorry there is no ready-to-run script.

However, I suggest to iterate through your dataset and process the image with the corruption type in corruptions/project/mmdet3d_plugin/corruptions.py. I guess it should not be that hard.

leesunmyung commented 5 days ago

Could you let me know your 'conda list'?

I have packages about mmdetection like below, but there may be some conflict.

mmcls 0.17.0 pypi_0 pypi mmcv 2.0.0 pypi_0 pypi mmcv-full 1.3.13 pypi_0 pypi mmdet 3.0.0 pypi_0 pypi mmdet3d 1.4.0 pypi_0 pypi mmengine 0.10.4 pypi_0 pypi mmsegmentation 1.0.0 pypi_0 pypi

I got some error like 'ImportError: canno import name ~ from ~', when bash tools/generate_dataset.sh

Daniel-xsy commented 5 days ago

Hi, the mmcv package is as follow:

mmcv-full                 1.4.0                    pypi_0    pypi
mmdet                     2.14.0                   pypi_0    pypi
mmdet3d                   0.17.1                   pypi_0    pypi
mmsegmentation            0.14.1                   pypi_0    pypi

Hope it can help. It would be helpful to debug if you could provide full error information. Another way is to refer to the mmdetection3d codebase for incompatibility issue.