Skip to content

README.md

Latest commit

 

History

History
98 lines (73 loc) · 2.98 KB

README.md

File metadata and controls

98 lines (73 loc) · 2.98 KB

Preprocess Scripts

This folder contains dataset-specific scripts that convert raw datasets to MotionCrafter training format:

  • video clips: *_rgb.mp4
  • annotations: *_data.hdf5
  • index file: meta_infos.txt

0) Unified launcher (recommended)

Use one command to run a dataset-specific preprocess script:

python run_preprocess.py \
  --dataset spring \
  --data-dir /path/to/raw/Spring \
  --output-dir /path/to/unnormed/Spring_video \
  --split train \
  --clip-length 150

For scripts using multi-splits:

python run_preprocess.py \
  --dataset gta_sfm \
  --data-dir /path/to/raw/GTA-SfM \
  --output-dir /path/to/unnormed/GTA-SfM_video \
  --splits train,test

1) Common configuration via environment variables

All gen_*.py scripts now support the same environment variables:

  • MOTIONCRAFTER_DATA_DIR: raw dataset root directory
  • MOTIONCRAFTER_OUTPUT_DIR: output directory for converted videos/hdf5
  • MOTIONCRAFTER_CLIP_LENGTH: frames per clip

Optional variables used by some scripts:

  • MOTIONCRAFTER_SPLITS: comma-separated split list (e.g. train,test)
  • MOTIONCRAFTER_SPLIT: single split name
  • MOTIONCRAFTER_PROCESS_SCENE_FLOW: true/false (Point Odyssey)

Example

cd datasets/preprocess
MOTIONCRAFTER_DATA_DIR=/path/to/raw/Spring \
MOTIONCRAFTER_OUTPUT_DIR=/path/to/unnormed/Spring_video \
MOTIONCRAFTER_SPLIT=train \
MOTIONCRAFTER_CLIP_LENGTH=150 \
python gen_spring_video.py

2) Normalize generated datasets

normalize_video_dataset.py supports CLI arguments:

python normalize_video_dataset.py \
  --data-dirs /path/to/unnormed/Spring_video /path/to/unnormed/GTA-SfM_video \
  --output-root /path/to/tmp_datasets \
  --resolution 320 640 \
  --num-workers 8 \
  --skip-existing

If --output-root and --output-dirs are not provided, output defaults to:

  • replace unnormed_datasets with tmp_datasets in data_dir, or
  • data_dir/normalized when replacement is not possible.

3) Build latent meta file

preprocess_meta_file.py generates latent meta_infos.txt from source meta files:

python preprocess_meta_file.py \
  --data-dirs /path/to/data_normed_1/Spring_video /path/to/data_normed_1/GTA-SfM_video \
  --latent-dirs /path/to/latent/Spring /path/to/latent/GTA-SfM

4) Notes

  • Most scripts assume CUDA is available for point map / flow computation.
  • Keep clip length consistent with your training setup.
  • Some datasets require dataset-specific directory structures exactly as expected by each script.
  • Common helper functions for env parsing and meta_infos.txt writing are in preprocess_common.py.

5) Dataset-specific dependencies

  • Dynamic Replica: requires pytorch3d.
  • Virtual KITTI 2: requires pandas.
  • Scripts reading EXR files (e.g. IRS/Matrix City): require OpenEXR runtime support.

If pip install -r requirements.txt cannot install pytorch3d in your environment, install it separately by following the official PyTorch3D installation guide that matches your PyTorch and CUDA versions.