PROJECT TITLE: PACKAGE DETECTION MODEL
TEAM MEMBERS:
• Janhavi Gupta
• Abhishek Gothe
• Vinit Badiger
• Adarsh Pandey
• Shahanwaj Idrisi
- USE CASE IMPORTANCE
Detecting packages and parcels in real time helps improve warehouse management, reduce manual errors, and optimize sorting operations in logistics. Our solution uses computer vision technology to identify packages on both mobile and desktop devices for faster and more efficient parcel tracking.
- DATA COLLECTION AND ANNOTATION
We utilized a robust dataset to achieve high accuracy in package detection.
Total Images: 6,377
Class: Package / Parcel
Dataset Split:
• Training Images: 4,464
• Validation Images: 1,275
• Test Images: 638
Data Augmentations Applied:
• Horizontal Flipping
• Brightness Adjustment
• Scaling
• Cropping
These augmentations helped the model handle different lighting conditions and complex backgrounds effectively.
- MODEL TRAINING AND VALIDATION
We trained the YOLOX model for object detection and monitored several important evaluation metrics.
Metrics Used:
• Precision
• Recall
• mAP@0.5
• Loss Values (Box Loss, Class Loss, DFL Loss)
Visual Analysis Tools:
• Confusion Matrix
• Precision-Recall Curve
• F1-Score Curve
These analyses helped us fine-tune and improve the model performance.
- MODEL DEPLOYMENT AND DEMO
The trained model was successfully deployed on the YOLOX mobile platform by WISERLI.
Key Features:
• Real-time package detection
• Satisfactory FPS performance
• Bounding boxes with confidence scores
Demo Video:
CONCLUSION AND LEARNINGS
The model performed well during live testing. Initially, lighting variations and complex backgrounds created challenges, but dataset augmentation significantly improved the model’s generalization capability.
This project provided valuable hands-on experience in:
• Dataset annotation consistency
• YOLOX optimization
• Real-time AI deployment
Special Thanks:
We sincerely thank Dr. Chandrakant Bothe Sir and WISERLI for conducting this internship workshop and providing us with the opportunity to gain practical knowledge in AI deployment.