Problem with renderer while training the framwork
Closed this issue · 8 comments
Hello @deguchihiroyuki!
Thank you for open sourcing this nice work.
When I train, I am getting below error.
Number of points at initialisation : 100000 [23/09 13:28:51]
Training progress: 0%| | 0/30000 [00:00<?, ?it/s]Traceback (most recent call last):
File "train_BlurAndEvent.py", line 273, in
training(lp.extract(args), op.extract(args), pp.extract(args), args.test_iterations, args.save_iterations, args.checkpoint_iterations, args.start_checkpoint, args.debug_from)
File "train_BlurAndEvent.py", line 111, in training
render_pkg = Nrender(viewpoint_cam, gaussians, pipe, background, viewpoint_cam.image_name)
File "/mnt/data/trisha/3dgs/E2GS/gaussian_renderer/init.py", line 159, in Nrender
debug=pipe.debug
TypeError: new() got an unexpected keyword argument 'viewmatrix'
It seems as per my understanding, rendered is slightly modified that from standard 3DGS renderer.
Can you please update renderer installation details or help me to resolve this error.
Thank you
Hi @trishagithubreddypalli , thank you for interested in my work!
I'm sorry but I can't get the reason why the error happens because I confirmed train_BlurAndEvent.py works with the code I uploaded on github.
I think 'viewmatrix' is the argument of the "class GaussianRasterizationSettings(NamedTuple):" in G:\ssd1\anaconda3\envs\gaussian_splatting\Lib\site-packages\diff_gaussian_rasterization
So, please check about this. And I use diff_gaussian_rasterization=0.0.0
Thank you so much for the help .The error is resolved .
When I am working with the synthetic dataset , after the training completion I am able to run the render.py file .But when I am using real world dataset , the training process is done perfectly but when I am running the file render.py I am getting error like
ERROR:
Looking for config file in /mnt/data/trisha/3dgs/e2gs/E2GS/output/7009da11-3/cfg_args
Config file found: /mnt/data/trisha/3dgs/e2gs/E2GS/output/7009da11-3/cfg_args
Rendering /mnt/data/trisha/3dgs/e2gs/E2GS/output/7009da11-3
Loading trained model at iteration 30000 [26/09 12:20:18]
Traceback (most recent call last):
File "/mnt/data/trisha/3dgs/e2gs/E2GS/render.py", line 123, in
render_sets(model.extract(args), args.iteration, pipeline.extract(args), args.skip_train, args.skip_test)
File "/mnt/data/trisha/3dgs/e2gs/E2GS/render.py", line 87, in render_sets
scene = Scene(dataset, gaussians, "no_blur_loss", load_iteration=iteration, shuffle=False)
File "/mnt/data/trisha/3dgs/e2gs/E2GS/scene/init.py", line 88, in init
self.cameras_extent = scene_info.nerf_normalization["radius"]
UnboundLocalError: local variable 'scene_info' referenced before assignment
Can you please help me out from these .Where can be the mistake?
Thank you for your time.
And also when I am reconstructing the scene using the colmap . I am getting some NANs in pointcloud.ply file .
Where will be the issue? I will be grateful if you help me out from this issue.
Thank you
If it is possible, can you please share the dataset after applying EDI model for deblurring.
Thank you!
Yes, I also tried but images are not deblurred properly, hence colmap is not working on these..
Please provide dataset. It would be much appreciable, if you can share the dataset which directly supports gaussian splatting(includes pose information as well sparse point cloud)..
Thank you so much
@trishagithubreddypalli @GopiRajuMatta , thanks for raising the discussion. I will upload the code to apply EDI model later.
And, I create the initial point cloud like the way below.
- Using "colmap gui"
- Doing "automatic reconstruction"
- After finishing it, "\dense\0\sparse\points3D.bin" shoould be made, and use it as an initial pointcloud.
Even if inputs are the E2NeRF's blurry images, I think it would work. So please try to do so!
UnboundLocalError: local variable 'scene_info' referenced before assignment
Hello @trishagithubreddypalli ! I have encountered a similar problem as well. It seems that the scene_info under no_blur_loss is not defined in the source code? Have you found a solution yet? And is it reasonable that the rendered result after I changed no_blur_loss to blur_loss and real_dender showed a much lower BRISQUE value than in the paper? @deguchihiroyuki