Object Detector Analysis

This repository contains code, data, and analysis for a comparative study of modern object detection algorithms: Faster R-CNN, YOLO-World, and GroundingDINO. The project evaluates models on their bounding box accuracy and prompt-based detection capabilities, using metrics like IoU, Precision, Recall, and F1 Score.

📄 View full project report (PDF)

---

🔍 Objective

To analyze the effectiveness of prompt-based and traditional object detectors across:

• Standard object detection (e.g., apples)
• Prompt-based open-vocabulary detection
• Robustness under diverse, real-world conditions

---

Models Analyzed

1. Faster R-CNN (ResNet-50)

• Closed-vocabulary detector pretrained on COCO
• Reliable for fixed-label detection
• Not suitable for prompt-based or novel object recognition

2. YOLO-World

• Single-stage detector with CLIP-based text embedding
• Supports open-vocabulary detection using prompts
• Highly sensitive to threshold tuning and prompt wording

3. GroundingDINO

• Transformer-based architecture using CLIP for text grounding
• Achieved highest IoU and F1 scores
• Robust to prompt rewording and image complexity

---

 
   ObjectDetectorAnalysis/ 
   ├── apl/ # Apple image dataset for evaluation 
   ├── data/ # Dataset for robustness testing 
   ├── ResultIMGS/ # Model prediction output images 
   ├── FinalModelComp.ipynb # Comparison of YOLO-World, RCNN, GroundingDINO 
   ├── WeirdImgData.ipynb # Robustness testing on challenging images 
   ├── README.md # Project overview and usage instructions 
   └── .DS_Store # (System file — safe to delete) 

---

Datasets

Apple Detection Dataset

Used to benchmark box prediction accuracy across models.

Robustness Evaluation Dataset

Curated to test:

• Complex scenes
• Small and overlapping objects
• Occlusions
• Unusual prompts
• Blurry / noisy images
• Symbol and scene detection

---

Metrics

• Intersection over Union (IoU)
• Precision / Recall
• F1 Score

Key Findings:

• GroundingDINO performed best in almost all categories.
• YOLO-World requires precise threshold tuning and prompt phrasing.
• Faster R-CNN lacks prompt support, good only for known COCO classes.

Model	Box Thresh	Text Thresh	Precision	Recall	F1 Score	Mean IoU
GroundingDINO	0.35	0.2	0.690	0.825	0.712	0.830
YOLO-World	0.08	-	0.673	0.910	0.732	0.897
Faster R-CNN	-	-	✓ (COCO only)	✗ (no prompts)	✗	✗

---