Open zkaiWu opened 1 year ago
almost the same.I use vanilla-nerf to train the fox dataset of instant-ngp but get awful result with blurry color blocks but no convergence. I use the same command "ns-process-data ..." and get a transforms.json and then use command "ns-train vanilla-nerf ...... nerfstudio-data" to train it. Also ,I notice that it works the same in mipnerf , and it's weird that its training time remain the same when I use different numbers of pictures to train.
The vanilla-nerf
and mipnerf
models are setup to work with the blender synthetic data, not real-world data. For real-world data, use nerfacto
or instant-ngp
和模型设置为使用搅拌机合成数据,而不是真实世界的数据
vanilla-nerf
。mipnerf
对于真实世界的数据,使用nerfacto
或instant-ngp
How can I prepare this Blender data to train and reconstruct on Mip-NeRF or Vanilla-NeRF by using colmap? I also noticed that if the command ns-train is used without the nerfstudio-data suffix, the system assumes the data format is Blender data, but it throws an error saying that transforms_train.json cannot be found.
The
vanilla-nerf
andmipnerf
models are setup to work with the blender synthetic data, not real-world data. For real-world data, usenerfacto
orinstant-ngp
Are there any mipnerf
parameters we could try to change to make it more "suitable" for real-world bounded data like this?
Hello, I would like to ask if you have solved this problem?
I use the command
"ns-process-data images --data /data5/wuzhongkai/data/dreamfusion_data/llff/nerf_llff_data/flower --output-dir /data5/wuzhongkai/data/dreamfusion_data/llff/nerf_llff_data/flower --skip_colmap --colmap_model_path sparse/0 --skip_image_processing"
to generate transform.json and use the command
CUDA_VISIBLE_DEVICES=$1 ns-train nerfacto --data /data/nerf_llff_data/flower/transforms.json \ --experiment-name llff/flower --vis wandb \ nerfstudio-data
to train the model, but get bad result. Are there any key points to train the nerfacto on llff