TQTQliu
commented
3 weeks ago Hi, indeed, as you said, pixelsplat does not provide dataloaders for the DTU, LLFF, nerf_synthetic, or T&T datasets, so we wrote the corresponding dataloader and then reproduced the metrics of pixelsplat. Attached are the corresponding dataset files. You can put the dataset file under pixelsplat/src/dataset. In addition, you need to write a config file for each dataset, please refer to pixelsplat/config/experiment/acid.yaml.
Here is an example of llff config:
# @package _global_
defaults:
- override /dataset: llff
- override /model/encoder: epipolar
- override /model/encoder/backbone: dino
- override /loss: [mse, lpips]
wandb:
name: llff
tags: [llff, 512x512]
dataset:
image_shape: [512, 512]
roots: [nerf_llff_data]
data_loader:
train:
batch_size: 1
trainer:
max_steps: 600_001
If you want to obtain the reproduced results in our paper, please refer to the evaluation file to calculate the metrics.
Please note that the dataset file we provided has not been cleaned, hopefully this will help you. datasets.zip
Hi, thank you for the experiments compared with other methods. However, I found that the original pixelsplat and new MVSplat did not provide dataloaders for LLFF, nerf_synthetic or T&T datasets. But you have these scores in your tables. Could you please kindly share the scripts you used to reproduce scores in such datasets? I am sure this will help the society a lot.