New tracking-enhanced prompt method wins MOSE Challenge by improving SAM3's ability to segment tiny and complex video objects

TEP (Tracking-Enhanced Prompts) achieves first place with 56.91% accuracy on the 5th PVUW MOSE Challenge by combining external tracking models with multimodal large language models

Addresses SAM3's key limitation: poor segmentation of tiny and semantic-dominated objects in cluttered video environments

Training-free approach leverages existing models rather than requiring new model training, making it more practical and accessible

Designed specifically for complex video object segmentation tasks that demand robust target comprehension and environmental adaptability