ksam98nyu/assignment-2-rav4-semantic-search

video_detections_50_epochs.parquet

This file is the detection corpus for Assignment 2: image-to-video semantic retrieval.

• File: video_detections_50_epochs.parquet
• Rows: 1100 (one row per detection)
• Source video: YcvECxtXoxQ
• Frame sampling interval: 5 seconds
• Detector: YOLO segmentation model fine-tuned for 50 epochs

Schema

Column	Type	Description
video_id	string	YouTube video id for the corpus video.
frame_index	int	Frame number from sampled frames (frame_0001.jpg, etc.).
timestamp_sec	int	Time offset in seconds for the sampled frame.
class_id	int	Numeric class id predicted by the detector.
class_label	string	Semantic part label (for retrieval matching).
confidence_score	float	Detector confidence for this detection.
bounding_box	list[float]	[x_min, y_min, x_max, y_max] coordinates in pixels.
x_min	float	Left x-coordinate of bounding box.
y_min	float	Top y-coordinate of bounding box.
x_max	float	Right x-coordinate of bounding box.
y_max	float	Bottom y-coordinate of bounding box.
frame_file	string	Frame filename in frames/ (e.g., frame_0123.jpg).
detector_name	string	Weights filename used for inference.

How It Was Generated

1. Fine-tune model with part_a_finetune_seg_model.py (50 epochs in this run).
2. Run part_b_retrieval_corpus.py over all sampled frames.
3. Save one row per detection into this parquet file.

Notes

• Retrieval in Part C uses class_label overlap between query detections and this corpus.
• Temporal grouping is done using timestamp_sec with a 5-second gap rule.