saks2002/rav4-video-detections

Assignment 2: Image-to-Video Semantic Retrieval

Dataset URL: https://huggingface.co/datasets/saks2002/rav4-video-detections

Assignment 2 Report: Image-to-Video Semantic Retrieval

1. Detector Choice and Configuration

For this assignment, I chose the YOLOWorld-l (Large) model for object detection.

Why YOLOWorld?

The assignment required detection at the "object part level" (e.g., hoods, wheels, headlights). While standard YOLOv8 models (trained on COCO) are excellent at detecting "cars," they do not natively provide labels for specific interior or exterior components.

YOLOWorld is an open-vocabulary object detector. It allows for zero-shot detection by specifying target classes at runtime. This enabled the system to identify 14 specific car components:

• car hood, car bumper, car door, car wheel, car headlight, car taillight, car windshield, car trunk, car side mirror, car grille, car roof, car door handle, car license plate.

Configuration

• Model: yolov8l-world.pt
• Threshold: 0.05 for video frames (maximizing recall for the index) and 0.15 for query images.
• Hardware: NVIDIA GeForce RTX 4070 Laptop GPU (CUDA 12.8).

2. Video Sampling Strategy

The input video is a 46-minute car exterior walkthrough (YcvECxtXoxQ).

• Sampling Rate: 1 frame every 5 seconds.
• Total Frames: 559 frames extracted using ffmpeg.
• Rationale: Sampling every 5 seconds provides a balance between temporal resolution and processing efficiency. Given the slow-moving nature of the camera in the walkthrough, most components remain visible for at least 5-10 seconds, ensuring they are captured in the index.

3. Image-to-Video Matching Logic

The retrieval system operates in three phases:

1. Query Analysis: The rav4-exterior-images dataset frames are passed through the same YOLOWorld-l detector. The detected classes (e.g., "car wheel") are extracted.
2. Index Lookup: The system queries the detections.parquet file for all frames containing the identified class.
3. Temporal Aggregation (Merging):
   • Timestamps of matching frames are sorted.
   • Contiguous timestamps are merged into "clips" if the gap between them is less than 15 seconds.
   • Intervals with at least 1 supporting detection are returned.
   • The output includes the start_timestamp, end_timestamp, and a youtube_verify_url with parameters to jump directly to the segment.

4. Failure Cases and Limitations

• Small Components: Tiny parts like car door handle had lower detection confidence and were sometimes missed in far-away shots.
• Contextual Vagueness: Labels like car door sometimes triggered on the entire side of the car, which is technically correct but less precise than a tight bounding box on a specific door.
• Occlusion: When the camera moves quickly or the part is partially obscured (e.g., a wheel behind a curb), the interval might be split into two smaller segments if the gap exceeds 15 seconds.
• Zero-Shot Bias: While YOLOWorld is powerful, its performance depends on the descriptive quality of the class names. Adding the "car " prefix significantly improved results over generic nouns like "hood."

5. Data Engineering

The results are stored in standardized Parquet files:

• detections.parquet: 1,321 detections across the video corpus.
• retrieval_results.parquet: Mapping for the 65 query images to retrieved video intervals.
• Accessible at: HuggingFace: saks2002/rav4-video-detections