Difference pose between gt and render images

[Gaaaavin]

I was able to train the model smoothly with my own data. But I have a problem when I tried to render the images from the trained model. I tried to run the following code:

python render_hierarchy.py -s ${DATASET_DIR} --model_path ${DATASET_DIR}/output --hierarchy ${DATASET_DIR}/output/merged.hier --out_dir ${DATASET_DIR}/output/renders --scaffold_file ${DATASET_DIR}/output/scaffold/point_cloud/iteration_30000

The is the same as that from the bottom of the readme, except that without the --eval flag. I changed this because I didn't use the --eval flag when training and I don't have the test.txt file.

However, I find that the rendered images and the gt images have a different pose. For example, the following two images are the first image in my data.

Please let me know what is the code that I should modify.

[Gaaaavin]

It turns out the problem was introduced by using the pose before reorientation (i.e. from the rectified folder, instead of the aligned folder). To successfully run the hierarchical rendering for the whole scene, you need to have images and sparse folder under DATASET_DIR, which doesn't exist if you follow the data preparation procedure. My solution is to create symbolic links to the folder. Go to your DATASET_DIR and run

ln -s camera_calibration/aligned/sparse/ sparse
ln -s camera_calibration/rectified/images/ images

This should solve the problem.