data

Pre-processed data

We provide pre-processed .h5 files for SURREAL and Mixamo dataset. You can find the processed .zip files here.

The data folder (i.e., data/) should be organized as the following:

├── data
│   ├── surreal
│   │   ├── surreal_train_h5py.h5
│   │   │
│   │   ├── surreal_val_h5py.h5 
│   │   │
│   │   └── surreal_val_idxs.npy # selected views/poses indices for evaluation 
│   │   
│   ├── mixamo 
│   │   ├── James_processed_h5py.h5
│   │   │
│   │   ├── James_selected.npy # selected indices used in the paper
│   │   │
│   │   ├── Archer_processed_h5py.h5
│   │   │
│   │   └── Archer_selected.npy # selected indices used in the paper

Note that Surreal dataset has ground truth pose and camera data, while Mixamo dataset only has SPIN estimated ones.

Pre-trained weights

We also provide pre-trained characters for SURREAL and Mixamo. You can download the models here.

For rendering, you can simply use the config files in configs/ for --nerf_args.

Belows are the brief summaries for the pretrained models:

surreal.tar: A model trained on SURREAL full dataset for 150k iterations with A-NeRF.

{james,archer}_ft.tar: Models that are trained with pose refinement for 500k (with --opt_pose_stop 200000), and then finetune with config file configs/mixamo/mixamo_finetunes.txt for 200k.

{james,archer}_ft_tv.tar: Same as the above, but use --use_temp_loss --temp_coef 0.05 during the pose refinement phase.