ntdky/video-detections

• Video Semantic Index Dataset

  Overview

  This dataset contains frame-level object detection results extracted from sampled video frames.

  Detections were generated using a YOLO-based object detection model and stored in Apache Parquet format for efficient structured indexing and downstream semantic retrieval.

  Each row represents a single detected object instance.

  ---

  File Information

  • Filename: video_semantic_index.parquet
  • Format: Apache Parquet
  • Compression: SNAPPY
  • Created with: pyarrow 18.1.0
  • Total rows: 1,398
  • Row groups: 1
  • Total compressed size: 59,376 bytes

  ---

  Schema Description

  Each row corresponds to a single object detection from a sampled video frame.

  Column Name	Type	Description
  video_id	string	Identifier of the source video
  frame_index	int64	Index of the sampled frame
  timestamp_sec	int64	Timestamp of the frame (in seconds)
  class_label	string	Detected object category label
  bounding_box	list[double]	Bounding box coordinates
  confidence_score	double	Detection confidence score

  ---

  Bounding Box Format

  Bounding boxes are stored as:

  [xmin, ymin, xmax, ymax]
  

  • Format: LIST logical type in Parquet
  • Each detection contains exactly 4 floating-point values
  • Total stored bounding box elements: 5,592

  Values represent pixel coordinates in the original frame resolution.

  ---

  Dataset Statistics

  Frame Information

  • Frame index range: 0 – 558
  • Total unique frames: 559

  Timestamp

  • Timestamp range: 0 – 2790 seconds

  Confidence Score

  • Range: 0.30 – 0.981

  Class Labels

  • Stored as UTF-8 strings
  • Dictionary-encoded for storage efficiency
  • No null values

  ---

  Storage Configuration

  • Encoding: PLAIN / RLE / RLE_DICTIONARY
  • Compression Codec: SNAPPY
  • Dictionary pages enabled for string columns
  • Single row group

  ---

  Data Generation Pipeline

  1. Input video processed offline.
  2. Frames sampled at fixed temporal intervals.
  3. YOLO-based object detection model applied per frame.
  4. Each detected object stored as one row.
  5. Bounding boxes serialized as list[float64].
  6. Data exported using PyArrow 18.1.0.
  7. SNAPPY compression applied.

  ---

  Notes

  • Each row represents a single detected object.
  • Multiple rows may correspond to the same frame.
  • No missing or null values are present.
  • This dataset serves as the frame-level semantic index, which can be further aggregated into temporal segment-level indices for retrieval tasks.