CoolFace
Datasetpublic

rushilara/h2-video-detections

Video Detections and Query Clips Parquet outputs for the video detection and image semantic search pipeline (car-parts detector on video + RAV4 query images). Files video_detections.parquet — One row per video frame; each row has a list of object detections for that frame. query_longest_clips.parquet — One row per query image; each row has the longest contiguous video clip where the detected car parts appear, with a YouTube embed URL. Schema… See the full description on the dataset page: https://huggingface.co/datasets/rushilara/h2-video-detections.

sourceHugging Faceupdated 7mo agoView on Hugging Face
0likes5downloads
Dataset Card

Video Detections and Query Clips

Parquet outputs for the video detection and image semantic search pipeline (car-parts detector on video + RAV4 query images).

Files

  • —video_detections.parquet — One row per video frame; each row has a list of object detections for that frame.
  • —query_longest_clips.parquet — One row per query image; each row has the longest contiguous video clip where the detected car parts appear, with a YouTube embed URL.

Schema

video_detections.parquet

ColumnTypeDescription
video_idstringIdentifier of the video (e.g. "input_video").
frame_indexintFrame number (1-based).
detectionslistList of detections for this frame. Each element has the fields below.

Each element in detections:

FieldTypeDescription
class_labelstringCar part class (e.g. "hood", "front_bumper", "wheel").
bounding_boxlist[x_min, y_min, x_max, y_max] in image coordinates.
confidence_scorefloatDetection confidence in [0, 1].

querylongestclips.parquet

ColumnTypeDescription
query_image_idintIndex of the query image in the dataset.
start_secintStart time of the longest contiguous clip (seconds).
end_secintEnd time of the longest contiguous clip (seconds).
youtube_embed_urlstringURL to embed the clip (e.g. https://www.youtube.com/embed/VIDEO_ID?start=...&end=...).