README.md

Assignment 2 – Image-to-Video Part Retrieval Dataset

This dataset contains the outputs of a detection-based image-to-video retrieval system.

It includes:

1. A video detection index (frame-level detections stored in Parquet format).
2. Retrieval results mapping query images to contiguous video segments.

---

Dataset Structure

1. video_detections.parquet

Frame-level detections extracted from a sampled video clip.

Each row corresponds to a single detection in a single frame.

Columns

• video_id: YouTube video ID
• frame_index: Index of the sampled frame
• timestamp_sec: Timestamp of the frame in seconds
• class_label: Predicted part label
• bounding_box: [x1, y1, x2, y2]
• confidence_score: Detector confidence

This file serves as the indexing layer for retrieval.

---

2. retrieval_results.parquet

Query-to-video retrieval results.

Each row corresponds to one retrieved contiguous time segment.

Columns

• image_id: Query image identifier
• video_id: YouTube video ID
• class_label: Matched part label
• start_timestamp: Segment start (inclusive)
• end_timestamp: Segment end (exclusive)
• number_of_supporting_detections: Number of frame-level detections supporting this segment