Sign in to view source links and access this dataset
Description
DREAM-1K is a challenging video description benchmark containing 1,000 short video clips. The videos are sourced from five diverse origins, including live-action movies, animated movies, stock videos, long YouTube videos, and TikTok-style short videos. Each video clip is manually annotated with fine-grained descriptions.
Use Cases
Train video captioning models based on fine-grained manual annotations.
Benchmark multimodal AI systems on diverse video sources mentioned in the description.
Evaluate model robustness across video complexities from different origins like movies and social media.
Study the relationship between visual events, actions, and their textual descriptions.
Strengths
Contains 1,000 video clips, providing a substantial benchmark scale.
Videos are sourced from five diverse origins, increasing content variety.
Each video has fine-grained manual annotations, ensuring high-quality labels.
Video clips are standardized to a short duration of around 10 seconds each.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
Last updated 2024-09-14 05:03:03; freshness should be verified.
Provenance
Source
huggingface
Collection Method
Manually annotated collection of video clips from five specified origins.
Freshness
Last updated 2024-09-14 05:03:03.
License is unknown; users must verify permissions before use.