MotionScape

A Large-Scale Real-World UAV-View Video Benchmark for Evaluating Future Visual Prediction in World Models.

[Benchmark Access]

---

Overview

MotionScape is a real-world UAV-view video benchmark designed for evaluating future visual continuation and visual prediction under highly dynamic aerial viewpoints. It contains diverse scenes and weather conditions, and is intended to support research on UAV visual prediction, future visual continuation, world model evaluation, spatiotemporal modeling, and text-conditioned video generation.

MotionScape provides clip-level captions that describe scene content, environmental conditions, and camera-motion tendencies. These captions serve as descriptive textual conditions for analyzing future visual generation under aerial motion, and are not intended to serve as explicit action commands or control signals.

Data Access

The benchmark is currently provided for academic research purposes only. At this stage, we release a partial subset of MotionScape for preview and review purposes. The complete benchmark and related resources will be released in a later version according to the project schedule.

• Current release: partial MotionScape data are available through the Hugging Face repository.
• Full release: the complete benchmark, metadata, evaluation splits, captions, and benchmark resources will be released in a future update.
• Access request: users can request access through the Hugging Face gated dataset page by reviewing and accepting the usage terms.
• Usage terms: access is granted for academic research use only and is subject to the benchmark license and terms of use.

More details about the release format, download instructions, benchmark protocol, and usage policy will be updated on this page.

License and Terms of Use

MotionScape is intended for academic research use only. Users must comply with the corresponding usage terms, citation requirements, and benchmark license. Detailed license information will be provided here.