The custom dataset reconstruction result is incorrect

[bigbigfo]

Hi, thank you very much for your excellent work. I am having some difficulty training my own dataset.My own dataset is an underwater dataset. I ran the tools/convert.py script you gave me on my dataset to generate the colmap result.
Because I am using the OPENCV camera model, an error occurs during training:
RuntimeError: Unsupported camera model: only PINHOLE or SIMPLE_PINHOLE currently. Please undistort your images with the command below first:
colmap image_undistorter --image_path data/sanya0/train/block_all/images_1.2 --input_path data/sanya0/train/block_all/sparse/0 --output_path data/sanya0/train/block_all/dense
then use data/sanya0/train/block_all/dense as the value of --data.path. I ran the colmap command according to the error message and was able to train normally.
My colmap results are correct, but my coarse training results seem incorrect.
I have a few questions:

1. Running tools/convert.py generates the undistorted image images. So when running the “colmap image_undistorter --image_path data/sanya0/train/block_all/images_1.2 --input_path data/sanya0/train/block_all/sparse/0 --output_path data/sanya0/train/block_all/dense” command, is the original image or the undistorted image used?
2. Is it possible that the incorrect training results are due to a problem with my colmap results being converted from the OPENCV camera model to PINHOLE?Could you please tell me how to correctly convert the colmap result from OPENCV camera model to PINHOLE model?
3. Are there other reasons why my training results are incorrect, and are there any solutions?

Looking forward to your reply！

[DekuLiuTesla]

@bigbigfo Is there any result on the test view?

[bigbigfo]

@DekuLiuTesla Sorry, I just saw your message. It may be because the amount of my data is too small, resulting in inaccurate results of the rough training. The results got better after I did the partition training.
I want to know now, firstly running tools/convert.py to generate sparse/0 and the undistorted images and then use the undistorted images to run the command "colmap image_undistorter --image_path data/sanya0/train/block_all/images --input_path data/sanya0/train/block_all/sparse/0 --output_path data/sanya0/train/block_all/dense" to convert the OPENCV camera model to a PINHOLE camera model ,is it correct?