Dataset Preparation

March 12, 2026 · View on GitHub

To train MotionCrafter, please download the following datasets from their official project pages.
Note that the paths such as data_normed_1/... refer to preprocessed versions used in our implementation.
You should first download the raw datasets listed below and then run our preprocessing pipeline to obtain the normalized format.

DynamicReplica

Replica synthetic indoor dataset used in dynamic scene reconstruction research.
Project Page: https://github.com/facebookresearch/Replica-Dataset

GTA-SfM

Synthetic outdoor dataset rendered from GTA for SfM, depth, and motion estimation tasks.
Reference: https://github.com/HKUST-Aerial-Robotics/Flow-Motion-Depth

Kubric

Large-scale synthetic multi-task dataset generated using the Kubric engine from Google Research.
Project Page: https://github.com/google-research/kubric

MatrixCity

Ultra-large-scale city-level synthetic dataset for driving and urban scene perception.
Project Page: https://github.com/city-super/MatrixCity

MVS-Synth (GTAV_720)

Multi-view stereo synthetic dataset rendered from GTA-V.
Dataset Page: https://phuang17.github.io/DeepMVS/mvs-synth.html

Spring

High-quality synthetic dataset for optical flow, scene flow, stereo, and multi-task perception.
Benchmark Website: https://spring-benchmark.org/

Point Odyssey

Large synthetic long-term point tracking dataset.
Project Page: https://pointodyssey.com/

Synthia

Synthetic urban driving dataset widely used for segmentation and depth estimation.
Official Page: https://synthia-dataset.net/

TartanAir

AirSim-based large-scale SLAM and multi-modal perception dataset.
Dataset Page: https://tartanair.org/

Virtual KITTI 2

Photorealistic synthetic clone of KITTI for driving perception tasks.
Dataset Page: https://europe.naverlabs.com/Research/Computer-Vision/Proxy-Virtual-Worlds/

BlinkVision

Video + event-based multimodal synthetic benchmark.
Website: https://www.blinkvision.org/

OmniWorld

Multi-domain, multi-modal synthetic world model dataset.
Project Page: https://yangzhou24.github.io/OmniWorld/

ScanNet++

High-resolution indoor real-world 3D reconstruction dataset.
Project Website: https://scannetpp.mlsg.cit.tum.de/scannetpp/

📝 Notes

  • The datasets above provide raw data. To reproduce the training pipeline, run the preprocessing script to convert them into the data_normed_1/... structure.
  • Several datasets are large (100–500K frames); please ensure sufficient storage.
  • Open-source preprocessing usage and unified script configuration are documented in datasets/preprocess/README.md.