MOSE: Video Object Segmentation in Complex Scenes from ICCV 2023

MOSE is a video object segmentation benchmark introduced at ICCV 2023 by Henghui Ding for tracking and segmenting objects in challenging environments. The dataset provides pixel-level masks for objects within video sequences characterized by heavy occlusions and crowded scenes.

Use Cases

• Training video object segmentation models using pixel-level masks
• Benchmarking tracking robustness against object disappearances and reappearances
• Evaluating temporal consistency of masks across sequential video frames

Strengths

• Peer-reviewed at ICCV 2023
• Focuses on complex environments with heavy occlusions
• MIT licensed for open research

Limitations

• High computational overhead typical of high-resolution video segmentation tasks
• Metadata regarding specific object classes is not detailed in the repository summary

Provenance

Source

ICCV 2023 publication by Henghui Ding

Collection Method

Annotated video sequences

Freshness

Last updated September 2025; based on 2023 research.