rakshi719/GOT10k-I2V

GOT-10k Image-to-Video Retrieval

MTEB/MOEB representation of the GOT-10k validation split for image-to-video retrieval.

Task

Given the first frame of a tracking sequence, retrieve the full tracking video. The mapping is one-to-one: each query has exactly one relevant item (the other direction of the same sequence).

Contents

• Queries: 180 first-frame images
• Corpus: 180 tracking videos
• Qrels: 180 one-to-one binary relevance judgments

Source

GOT-10k (Generic Object Tracking benchmark) validation split — 180 real-world sequences spanning 563 object classes and 87 motion patterns. Videos are encoded from the official JPEG frames at 10 fps with libx264.

Paper: https://arxiv.org/abs/1808.00803 Official site: http://got-10k.aitestunion.com/

License

CC-BY-4.0. See the official GOT-10k page for details.

Citation

@article{huang2019got,
  author = {Lianghua Huang and Xin Zhao and Kaiqi Huang},
  title = {GOT-10k: A Large High-Diversity Benchmark for Generic Object
           Tracking in the Wild},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year = {2019},
}