Hello,
thanks for your amazing work. I'm using your repo and the paper for textured mesh generation of custom objects for an object detection pipeline. I build lots of meshes and got some amazing results with the default settings:
I did some hyperparameter experiments for different objects to see how the results are influenced.
I had a similar problem for the most objects when changing the iterations in stage zero to a higher value (15000 and upwards): At some epoch the training fails and the training error is NaN. Here is a screenshot and the corresponding log file for stage zero with 30000 iterations (the rest is default setting for single object reconstruction)
Here is a log file from stage 0 for the elephant i showed above, when increasing the iterations to 30000:
log_ngp_stage0.txt
I used the recommended default setting for both stages for single object reconstruction in --sdf mode.
I saw the same behaviour for different objects at some point where i increased the number of iterations. For some objects (esp. with some shiny surfaces) I even saw this for the default setting.
All datasets were created using COLMAP, The sparse point cloud and the camera location are reasonable.
Did someone noticed a similar behavior or can explain this?
I am grateful for any input or tips and can provide additional input and results
Hello, thanks for your amazing work. I'm using your repo and the paper for textured mesh generation of custom objects for an object detection pipeline. I build lots of meshes and got some amazing results with the default settings:
Results for default setting
Here is an example of casual generated from casual smartphone images: nerf stage 1: https://github.com/ashawkey/nerf2mesh/assets/108192596/585dba7d-cc92-4689-9b84-2bc3d3269518 mesh:
Obeservations an Problem
I did some hyperparameter experiments for different objects to see how the results are influenced. I had a similar problem for the most objects when changing the iterations in stage zero to a higher value (15000 and upwards): At some epoch the training fails and the training error is NaN. Here is a screenshot and the corresponding log file for stage zero with 30000 iterations (the rest is default setting for single object reconstruction)
Here is a log file from stage 0 for the elephant i showed above, when increasing the iterations to 30000: log_ngp_stage0.txt I used the recommended default setting for both stages for single object reconstruction in --sdf mode.
I saw the same behaviour for different objects at some point where i increased the number of iterations. For some objects (esp. with some shiny surfaces) I even saw this for the default setting. All datasets were created using COLMAP, The sparse point cloud and the camera location are reasonable.
Did someone noticed a similar behavior or can explain this? I am grateful for any input or tips and can provide additional input and results
keep up the grate work Paul