Logo Detection and Classification

Code for the CS 534: Computer Vision project at Rutgers University. Reproduced and extended on the results presented in the paper Open Set Logo Detection and Retrieval

---

Work Done

• Reproduced and improved the mean average precision (mAP) of open set logo detection from the original paper by 24 percentage points using YOLOv5 detector
• Obtained a classification accuracy of 22.56% for 47 classes of the Flickr-47 dataset using a logo classification architecture consisting of YOLOv5 and template matching focused on both abstract and textual logos

---

Tech Stack

• Python 3
• PyTorch
• seaborn

---

Report and Presentation

You can access the report here and presentation here.

Structure and Acknowledgements

• utils: Contains code for file handling and bounding box annotation
• logo_classification.py: Contains code for the classification of logos in the Flickr-47 dataset
• YOLOv5_Custom_Training.ipynb: Contains application of YOLOv5 to the open set dataset and reports the improvement in mAP.

Made as a team with @animesharma