OmniLife360 is a benchmark for 3D reconstruction from in-the-wild 360-degree captures, containing 2,407 panoramic sequences totaling about 66 hours. The dataset, created by AutoLab-SJTU, includes 1,937 training and 470 testing sequences across 64 scene categories and 31 action classes. It provides source links, benchmark metadata, COLMAP/SfM reconstruction results, and helper scripts for frame preparation and panorama SfM reproduction.
Use Cases
- Benchmarking 3D reconstruction algorithms based on in-the-wild 360-degree video sequences.
- Training models for scene understanding based on 64 annotated scene categories.
- Developing action recognition systems based on 31 labeled action classes within panoramic video.
- Reproducing Structure-from-Motion (SfM) pipelines using the provided COLMAP results and helper scripts.
Strengths
- Contains 2,407 panoramic sequences providing a substantial scale for benchmarking.
- Offers about 66 hours of total video footage for training and analysis.
- Includes structured metadata with 64 scene categories and 31 action classes for detailed annotation.
Limitations
- Column-level documentation is absent; field semantics must be inferred after download.
- Row count is unknown, which may limit suitability assessment for specific modeling tasks.
Provenance
- Source
- AutoLab-SJTU via Hugging Face.
- Collection Method
- Likely collected as in-the-wild 360-degree captures; includes SfM reconstruction results.
- Time Range
- null
- Freshness
- Last updated 2026-06-26 12:14:50; freshness should be verified.
- Geography
- null