Bineet2407/rav4-video-detections

RAV4 Video Detections

Car-part detections and retrieval results from the Toyota RAV4 exterior review video (YcvECxtXoxQ) using a two-pass smart sampling strategy:

1. Coarse pass (1 frame / 5 s) scored by CLIP zero-shot exterior classifier (openai/clip-vit-base-patch32).
2. Dense pass (1 fps) applied only to windows with exterior confidence >= 0.70.
3. YOLOv8n-seg fine-tuned on the Ultralytics Car Parts Segmentation dataset runs on all dense frames.

---

File 1: video_detections.parquet

One row per bounding-box detection in the video. This is the searchable index — 6,653 detections spanning 0 s – 2,791 s.

Column	Type	Description
video_id	str	YouTube video ID (YcvECxtXoxQ)
frame_index	int	Timestamp rounded to nearest second
timestamp_sec	float	Exact time position in seconds
class_id	int	Detector class index (0–22)
class_label	str	Human-readable part name (e.g. wheel, hood)
x_min	float	Bounding box left edge (pixels)
y_min	float	Bounding box top edge (pixels)
x_max	float	Bounding box right edge (pixels)
y_max	float	Bounding box bottom edge (pixels)
confidence_score	float	Detector confidence [0, 1]
detector_name	str	Model identifier (yolov8n-seg-carparts)

---

File 2: retrieval_results.parquet

One row per retrieved video segment for a query image. 310 results across 65 query images from the aegean-ai/rav4-exterior-images dataset.

Column	Type	Description
query_index	int	Index of the query image in the dataset
query_timestamp	str	Source timestamp of the query image (MM:SS)
class_label	str	Car part class matched between query and video
start_timestamp	int	Segment start time in seconds
end_timestamp	int	Segment end time in seconds
number_of_supporting_detections	int	Frames within segment containing a detection of the matched class
youtube_link	str	YouTube embed URL with start/end parameters for direct verification