Deep Learning Object Detection Benchmark: YOLO-V5 VS Faster R-CNN on Pascal VOC 2012

Objective

This project implements and compares Faster R-CNN (two-stage detector) and YOLOv5 (single-stage detector) on the Pascal VOC 2012 dataset.

The goal is to analyze the trade-off between:

• Detection Accuracy (mAP)
• Inference Speed (FPS)
• Bounding Box Quality

---

Benchmark goal

Compare the two detector families on the same dataset (VOC2012):

• YOLOv5 (one-stage): optimized for speed/real-time detection.
• Faster R-CNN (two-stage): typically higher-quality detections, often slower inference.

---

Models Compared

1) YOLOv5 (One-stage Detector)

• Predicts bounding boxes + class probabilities in a single forward pass
• Typically:
  • Very fast inference
  • Strong baseline for real-time detection

2) Faster R-CNN (Two-stage Detector)

• Stage 1: Region Proposal Network (RPN)
• Stage 2: Classifies/refines proposals
• Typically:
  • Strong accuracy
  • Heavier compute and slower inference than YOLO-style models

---

Requirements

python>=3.7
torch>=1.9.0
torchvision>=0.10.0
yolov5
opencv-python
matplotlib
seaborn
scikit-learn
pillow
tqdm
pandas
numpy

---

Dataset

Pascal VOC 2012 (20 object classes).
This project was run on Kaggle Notebooks, so the dataset is expected to be attached to the Kaggle notebook session via “Add data”.

Typical Kaggle input location pattern:

• /kaggle/input/<dataset-name>/...

---

How to reproduce on Kaggle

1) Create a Kaggle Notebook

• Enable GPU (recommended).

2) Add dataset

• Add a Pascal VOC 2012 dataset through the Kaggle sidebar (“Add data”).
• Confirm the path by listing:

!ls /kaggle/input

3) Run YOLOv5 pipeline (from the repo notebook)

Open and run:

• YOLO V5/yolov5-execution.ipynb

What it produces (as committed here):

• YOLO V5/yolo_results_csv.csv
• Plot images under YOLO V5/yolo_results_plots/

4) Run Faster R-CNN pipeline (from the repo notebook)

Open and run:

• faster_rcnn_execution.ipynb

What it produces (as committed here):

• CSV files under rcnn_results_csv/
• Plot images under rcnn_results_plots/

---

Preparation

Pascal VOC 2012

• 20 object classes (person, car, bicycle, etc.)

• Dataset split:

  • Train: 70%
  • Validation: 20%
  • Test: 10%

Annotation Formats

• Faster R-CNN: Pascal VOC XML
• YOLOv5: YOLO TXT format

---

Models Implemented

🔹 Faster R-CNN (Two-Stage)

• Backbone: ResNet50 + FPN
• Stage 1: Region Proposal Network (RPN)
• Stage 2: Classification + Bounding Box Regression

🔹 YOLOv5 (Single-Stage)

• Real-time object detector

• Direct prediction of:

  • Bounding boxes
  • Class probabilities
  • Confidence scores

---

Training Pipeline

Faster R-CNN

• Optimizer: SGD

• Loss components:

  • Classification loss
  • Bounding box regression loss

• Epochs: 10

• Output:

  • Model weights (faster_rcnn_voc2012.pt)
  • CSV metrics
  • Detection visualizations

YOLOv5

• Pretrained weights: yolov5s.pt

• Epochs: ~100

• Automatic logging of:

  • Loss curves
  • mAP metrics
  • Validation performance

---

Faster R-CNN Pipeline

The Faster R-CNN workflow is provided primarily via:

• faster_rcnn_execution.ipynb

Typical Kaggle Setup

In the notebook, you generally need:

• PyTorch + torchvision (already available in many Kaggle GPU images)
• Common packages: numpy, pandas, matplotlib, opencv-python

YOLOv5 Pipeline

Environment Setup (Kaggle)

!python -V
!nvidia-smi

If you are using the official YOLOv5 repo approach:

!git clone https://github.com/ultralytics/yolov5
%cd yolov5
!pip install -r requirements.txt

Data Preparation (VOC → YOLO format)

YOLO expects:

• images in images/train, images/val
• labels in YOLO txt format in labels/train, labels/val
• a dataset YAML file like voc.yaml

You need a conversion step if your dataset is in VOC XML format.

Checklist:

• Confirm image paths
• Convert VOC XML annotations → YOLO txt labels
• Create voc.yaml with:
  • train: ...
  • val: ...
  • nc: 20
  • names: [ ...VOC class names... ]

Training

Example (adjust model size + epochs based on runtime budget):

!python train.py --img 640 --batch 16 --epochs 50 --data voc.yaml --weights yolov5s.pt

Inference

!python detect.py --weights runs/train/exp/weights/best.pt --img 640 --source <path-to-test-images>

Evaluation

YOLOv5 provides evaluation metrics (including mAP) via validation:

!python val.py --weights runs/train/exp/weights/best.pt --data voc.yaml --img 640

---

Training

The training loop typically includes:

• dataset loader (VOC)
• model initialization
• optimizer + scheduler
• epoch loop
• checkpoint saving

Recommended to log:

• training loss breakdown (RPN objectness, bbox regression, classifier loss, etc.)
• evaluation metric(s) each epoch (mAP if implemented)
• wall-clock time per epoch

Inference

Inference typically:

• loads model weights/checkpoint
• runs prediction on validation/test images
• applies confidence thresholding
• optionally applies NMS (usually handled internally)
• renders predicted boxes on images

Evaluation

For Faster R-CNN evaluation on VOC, you may implement:

• VOC mAP calculation (VOC07 11-point or VOC2010+ style)
• or COCO-style mAP if you convert annotations to COCO

Your repo indicates you save:

• CSV metrics → rcnn_results_csv/
• plots → rcnn_results_plots/

---

Notebooks (main execution)

• YOLOv5 notebook: YOLO V5/yolov5-execution.ipynb
• Faster R-CNN notebook: faster_rcnn_execution.ipynb

Outputs already committed (results)

YOLOv5 outputs

• Results CSV: YOLO V5/yolo_results_csv.csv
• Plots: YOLO V5/yolo_results_plots/
  • YOLO V5/yolo_results_plots/detection_results.png
  • YOLO V5/yolo_results_plots/training_metrics_curves.png
  • YOLO V5/yolo_results_plots/training_metrics_curves_2.png

Faster R-CNN outputs

• CSVs: rcnn_results_csv/
  • rcnn_results_csv/faster_rcnn_results.csv
  • rcnn_results_csv/faster_rcnn_training_metrics.csv
  • rcnn_results_csv/faster_rcnn_inference_speed.csv
  • rcnn_results_csv/faster_rcnn_detections.csv
  • rcnn_results_csv/faster_rcnn_class_distribution.csv
• Plots: rcnn_results_plots/
  • rcnn_results_plots/training_loss.png
  • rcnn_results_plots/training_metrics_curves.png
  • rcnn_results_plots/class_distribution.png
  • rcnn_results_plots/confidence_distribution.png
  • rcnn_results_plots/sample_detections.png
  • rcnn_results_plots/test_set_predictions.png

---

Performance Evaluation Metrics

1. Mean Average Precision (mAP)

• mAP@0.5 (VOC standard)
• mAP@0.5:0.95 (COCO-style)

2. Inference Speed

• Time per image
• Frames Per Second (FPS)

3. Detection Quality

• False positives / negatives
• Bounding box precision
• Confidence scores

---

Results (where to look)

YOLOv5

• Numeric summary: YOLO V5/yolo_results_csv.csv
• Training curves + detection visualization:
  • YOLO V5/yolo_results_plots/training_metrics_curves.png
  • YOLO V5/yolo_results_plots/training_metrics_curves_2.png
  • YOLO V5/yolo_results_plots/detection_results.png

Faster R-CNN

• Final metrics summary: rcnn_results_csv/faster_rcnn_results.csv
• Training metrics tracking: rcnn_results_csv/faster_rcnn_training_metrics.csv
• Inference speed logging: rcnn_results_csv/faster_rcnn_inference_speed.csv
• Detection-level outputs: rcnn_results_csv/faster_rcnn_detections.csv
• Visualizations:
  • rcnn_results_plots/training_loss.png
  • rcnn_results_plots/training_metrics_curves.png
  • rcnn_results_plots/class_distribution.png
  • rcnn_results_plots/confidence_distribution.png
  • rcnn_results_plots/sample_detections.png
  • rcnn_results_plots/test_set_predictions.png

---

Notes on fair comparison

To ensure YOLOv5 vs Faster R-CNN comparison is meaningful:

• Keep consistent dataset split (train/val/test) and class mapping (VOC 20 classes).
• Use clearly stated evaluation criteria (e.g., IoU threshold(s), confidence thresholds).
• Report timing with warm-up (first inference often slower).
• Mention the Kaggle GPU type used if available (affects speed).

---

Analysis

🔹 Faster R-CNN Performance

Training Loss Curve

• Smooth convergence from 0.39 → 0.15
• Stable learning without overfitting

---

📊 mAP Scores

• mAP@0.5: 0.6569
• mAP@0.5:0.95: 0.6569

---

Class Distribution of Detections

• Strong bias toward person class (1512 detections)
• Lower detection counts for rare classes (train, aeroplane)

---

Confidence Score Analysis

• Mean confidence: 0.885
• Most predictions fall in 0.9–1.0 range
• Indicates high model certainty

---

🔹 Sample Predictions (Faster R-CNN)

• Accurate localization in most cases
• Handles multiple objects well
• Some overlapping detections in dense scenes

🔹 Sample Predictions (YOLO V5)

---

🔹 Test Set Predictions (Faster R-CNN)

🔹 Test Set Predictions (YOLO V5)

• Strong performance in:

  • Multi-object scenes
  • Occlusion handling

• Minor false positives in crowded areas

---

YOLOv5 Results

Training Loss

• Smooth decreasing trend
• Faster convergence than Faster R-CNN

---

mAP Performance

• mAP@0.5: ~0.73
• mAP@0.5:0.95: ~0.60

---

Key Observations

• Faster inference (real-time capable)
• Slightly lower localization precision vs Faster R-CNN
• Better scalability for deployment

---

Final Comparison

Metric	Faster R-CNN	YOLOv5
mAP@0.5	0.6569	~0.73
mAP@0.5:0.95	0.6569	~0.60
Inference Speed	Slow	Fast
Localization	High Precision	Moderate
Real-time Capability	No	Yes

---

Key Insights

• Faster R-CNN

  • Better for accuracy-focused tasks
  • Strong bounding box precision
  • Suitable for research & analysis

• YOLOv5

  • Best for real-time applications
  • Faster and lightweight
  • Slight trade-off in precision

---

Conclusion

This project demonstrates the accuracy vs speed trade-off in object detection:

• Use Faster R-CNN when precision is critical
• Use YOLOv5 when speed is essential

---

Future Improvements

• Implement Faster-RCNN model
• Add quantization for mobile deployment
• Include confidence threshold analysis
• Add inference time comparison
• Implement model ensemble
• Add data augmentation experiments

References

• YoloV5 GitHub
• Faster-RCNN Paper
• PyTorch Documentation

Citations

• Pascal VOC: http://host.robots.ox.ac.uk/pascal/VOC/
• YOLOv5 (Ultralytics): https://github.com/ultralytics/yolov5
• Torchvision detection models (Faster R-CNN): https://pytorch.org/vision/stable/models.html

Author

Gubba Sai Ananya