TejaAlapati/assignment2-video-detections

Assignment 2 – Image-to-Video Semantic Retrieval

Overview

This dataset contains object detection results extracted from the YouTube video:

https://www.youtube.com/watch?v=YcvECxtXoxQ

The goal of this assignment was to build a semantic retrieval system that can identify video segments where a specific car exterior component appears.

Instead of manually labeling timestamps, the system detects objects in sampled video frames and stores structured detection results. These detections are later used for image-to-video retrieval.

Video Processing

The video was downloaded using yt-dlp and processed locally.

Frames were sampled at a rate of 1 frame every 5 seconds using the following command:

This sampling strategy reduces computation while still capturing meaningful content across the video timeline.

---

Object Detection

Object detection was performed using:

• Model: Ultralytics YOLOv8 (yolov8n.pt)
• Detection was applied to each sampled frame.
• No manual annotations or hard-coded timestamps were used.

Each detection generates one row in the dataset.

---

Dataset Schema

Each row in the Parquet file represents a single detection and includes:

• video_id (string): YouTube video ID
• timestamp (float): Timestamp in seconds corresponding to the sampled frame
• class_label (string): Detected object class
• x_min (float): Left bounding box coordinate
• y_min (float): Top bounding box coordinate
• x_max (float): Right bounding box coordinate
• y_max (float): Bottom bounding box coordinate
• confidence_score (float): Detection confidence score

---

Retrieval Method (Summary)

The retrieval system works as follows:

1. A query image is processed using the same detector.
2. The detected object class from the query image is identified.
3. Matching detections are retrieved from this dataset.
4. Consecutive timestamps are grouped into contiguous segments.
5. The system returns start and end timestamps for each segment.

---

Limitations

• Detection quality depends on the pretrained YOLOv8 model.
• Sampling every 5 seconds may miss very short object appearances.
• No temporal tracking or smoothing was applied.

This dataset serves as the structured detection layer required for semantic image-to-video retrieval.