ntdky/retrieval-results

• Semantic Retrieval Index Dataset

  Overview

  This dataset contains temporally aggregated semantic retrieval annotations extracted from video data. The file is stored in Apache Parquet format and optimized for efficient structured retrieval and indexing.

  Each row represents a semantic video segment candidate, aggregated from frame-level object detection results.

  ---

  File Information

  • Format: Apache Parquet
  • Compression: SNAPPY
  • Created with: pyarrow 18.1.0
  • Total rows: 101
  • Row groups: 1

  ---

  Schema Description

  Each row corresponds to a semantically indexed temporal video segment.

  Column Name	Type	Description
  start_timestamp	int64	Start timestamp of the candidate segment (in seconds)
  end_timestamp	int64	End timestamp of the candidate segment (in seconds)
  class_label	list[string]	List of detected semantic object labels associated with the segment
  number_of_supporting_detections	int64	Number of supporting detections aggregated into this segment
  youtube_link	string	Embedded YouTube URL pointing to the segment
  query_index	int64	Query identifier index used for retrieval evaluation

  ---

  Field Format Details

  Class Label Format

  The class_label column is stored as a Parquet LIST logical type.

  Example value:

  ["wheel", "back_bumper"]
  

  Each list contains the unique semantic categories detected within the temporal segment.

  ---

  Timestamp Statistics

  • start_timestamp range: 30 – 2455 seconds
  • end_timestamp range: 35 – 2460 seconds

  All timestamps are measured in seconds.

  ---

  Detection Statistics

  • number_of_supporting_detections range: 1 – 12

  This value represents how many frame-level detections were aggregated into the temporal segment.

  ---

  Query Index

  • query_index range: 9 – 62
  • Each query index corresponds to a retrieval query identifier used in evaluation.

  ---

  Storage Configuration

  • Encoding: PLAIN / RLE / RLE_DICTIONARY
  • Compression Codec: SNAPPY
  • Dictionary Pages: Enabled for optimized string storage
  • Total compressed size: 2,454 bytes

  ---

  Data Usage Notes

  • Each row represents an aggregated semantic video segment.
  • The dataset is designed for:
    • Query-to-video segment matching
    • Object-based semantic video retrieval
    • Detection-supported ranking evaluation
  • No null values are present in the dataset.
  • All fields are fully populated.

  ---

  Data Generation Pipeline

  1. Video corpus processed offline.
  2. Object detection applied at the frame level.
  3. Frame-level detections aggregated into temporal segments.
  4. Unique class labels collected per segment.
  5. Supporting detection counts computed.
  6. Structured annotations serialized into Parquet format.
  7. SNAPPY compression applied.