Naman02/rav4-exterior-retrieval-interval2

Assignment 2 — RAV4 Exterior Component Retrieval

This Hugging Face dataset repo contains the outputs for Assignment 2 (AI Spring 2026): a part-level object detector is used to index a YouTube video, and query images are used to retrieve relevant video segments.
This submission focuses only on exterior car components and indexes only the interval 18:43 to 24:43 of the original video (to reduce the number of frames analyzed).

Source video

• YouTube video ID: YcvECxtXoxQ
• Original video URL: https://www.youtube.com/watch?v=YcvECxtXoxQ
• Indexed interval (original timestamps): 18:43–24:43
  • start_sec = 1123, end_sec = 1483

Frame sampling

• Frames were sampled at 1 frame every 5 seconds from the indexed interval.
• frame_id is a 0-based index over sampled frames in the interval.
• timestamp_sec stored in detections is always in original YouTube time (i.e., timestamp_sec = 1123 + frame_id * 5).

---

Files

1) detections_by_frame.parquet — Detection Parquet (frame-level)

One row per frame. Multiple detections within the same frame are stored as a list in a single column.

Schema

• frame_id (int): 0-based sampled frame index within the interval.
• detections (list): list of detection objects for that frame. Each detection is a dict with:
  • class_label (string): exterior part label predicted by the detector (e.g., front_bumper, hood, left_mirror, front_light, etc.)
  • bounding_box (list[float]): [x_min, y_min, x_max, y_max]
  • confidence_score (float): detector confidence
  • timestamp_sec (int): detection timestamp in original YouTube seconds

---

2) clips_by_query.parquet — Clips/Segments Parquet (query → YouTube URL)

Two columns only, mapping each query to 1 or more retrieved YouTube clip URL (begin/end times in original video time).

Schema

• query_image (string): query identifier, formatted as query_<index>:<class_label>
• youtube_url (string): YouTube embed URL with begin/end seconds in original time:
  https://www.youtube.com/embed/YcvECxtXoxQ?start=<start_sec>&end=<end_sec>

3) best_clips_by_query.parquet — Best clip per query (one Youtube URL per query image)

one query image corresponds to exactly one best YouTube URL.

Schema

• query_image (string): identifier formatted as query_<index>:<class_label_used_for_retrieval>
• youtube_url (string): single best YouTube embed URL for that query, with begin/end seconds in original time.

How the “best clip” is selected (best_clips_by_query.parquet)

For each query image:

1. We run the part-level detector on the query image to obtain one or more predicted exterior class_label values.
2. For each predicted class_label, collect all matching detections from the indexed video interval (from detections_by_frame.parquet).
3. Then we convert the detection timestamps into contiguous time segments by merging detections that are at most 5 seconds apart (the frame sampling interval).
4. Then we score each candidate segment and choose the single best segment for the query. The score favors segments with:
   • more supporting detections within the segment (primary factor),
   • higher average confidence (tie-breaker),
   • longer duration (minor tie-breaker).
5. Then we output one row per query image containing the chosen class_label (embedded in query_image) and the corresponding YouTube embed URL with start/end in original video time.

---

Detector

• A YOLO segmentation model was fine-tuned on Ultralytics Car Parts Segmentation Dataset (carparts-seg).

Reproducibility notes

• The detection Parquet (detections_by_frame.parquet) is sufficient to reproduce the clip segments by:
  1. filtering detections by exterior class_label,
  2. grouping detections over time into contiguous segments,
  3. formatting segments as YouTube embed URLs with start/end