How to train my own datasets? about monoscene

[anhquancao]

Hi @lulianLiu,
You can follow these steps to write the dataloader for your dataset:

1. Adapt this file to generate 3D ground truth at 1:8 resolution
2. Modile kitti dataset loader with your camera intrinsic parameter. If you predict the 3D scene in the camera coordinate then T_velo_2_cam can be set to np.eye(4)
3. You might need to modify voxel_origin and scene size. This define the area of the scene you want to predict

[lulianLiu]

Hi @lulianLiu, You can follow these steps to write the dataloader for your dataset:

1. Adapt [this file](https://github.com/cv-rits/MonoScene/blob/master/monoscene/data/semantic_kitti/preprocess.py) to generate 3D ground truth at 1:8 resolution

2. Modile [kitti dataset loader](https://github.com/cv-rits/MonoScene/blob/master/monoscene/data/semantic_kitti/kitti_dataset.py#L69) with your camera intrinsic parameter. If you predict the 3D scene in the camera coordinate then T_velo_2_cam can be set to np.eye(4)

3. You might need to modify [voxel_origin and scene size](https://github.com/cv-rits/MonoScene/blob/master/monoscene/data/semantic_kitti/kitti_dataset.py#L41). This define the area of the scene you want to predict

Dear author, if I train by the kitti data. I need to download follow two data?Can I download the kitti(700M) to train and modile the data to make my own data?

[anhquancao]

You need the calibaration files and the RGB images which are not in the 700Mb data.

[lulianLiu]

But the RGB images is too big. Can you provide a small data link to me? and I also want to ask whether only use RGB images to train?

[anhquancao]

Sorry, I don't have the smaller dataset. You need to calibration file for the camera intrinsics and 3D ground truth to train

[AgentEXPL]

Hi, @anhquancao How to obtain the .pkl and .bin files from the original NYU dataset?