Preprocessing another data

[tuclen-3]

Hi @Owen-Liuyuxuan, Thanks for your great work, I have a problem when I inference my custom data. My model use your code to run KITTI data quite well but my custom data have sharp image quite different from the KITTI dataset (1280, 1980). I can see in your code have processing data code in the ./VisualDet3D/kitti/data directory but I don't know how to use it to process my custom data. Can you give me some advice to use it? Thanks

[myfun-deep]

Hello! I have the same problem as you . My data size is 720 1280. I modified the data.rgb_ shape, change it to 720 1280 3, but it doesn't work. I have another question, The size of KITTI dataset is 384 1280, but in the code, the rgb_shape is 28812803?

[tuclen-3]

Hi @myfun-deep, Did you fix your calibration when you modified image?

[myfun-deep]

calibration? I created my own dataset in the form of kitti.

[tuclen-3]

Hi @myfun-deep, I am just starting out in 3D field but if i guess i am not wrong then when we resize the image the intrinsic matrix will also change because the intrinsic matrix has 2 parameters as an optical center point that equai half length and width of image. Of course, when you resize image, that optical center change too. And we can't use old calibration to eval our new image. I think so. If I have wrong, please let me know. But in code of @Owen-Liuyuxuan have Preprocess class. I can see 2 process that CropTop and Normalize. In CropTop have an parameter is image_gt, I don't know what it is because Croptop with size is 100 will change image and don't change calib because I don't have image_gt and Normalize will make my image resize become black I can't see anything from my image.

[myfun-deep]

Hi @tuclen-3 I agree with you！But I don't know how to use my own data to replace KITTI. How to use my own data to train it

[Owen-Liuyuxuan]

The problem is formulating how you would like to reshape your image for training/testing.

For the KITTI dataset, the idea is to crop the top 100 pixels and then resize to 288 1280. Or some methods will directly resize the image to 384 1280.

For the design logic, most of you are correct that all input images and calibration matrix will be adjusted based on how we deal with the main images.

For 1920 * 1280, you can directly resize images to the shape you want, and the code should run fine.

[tuclen-3]

Hi @Owen-Liuyuxuan, So, i can use your Preprocess class with 2 processes are ConvertFloat and Resize (don't use CropTop and Normalize class). Right?