This repository contains the final group project of the team PACO for the course 20600 DEEP LEARNING FOR COMPUTER VISION taught by Gaia Rubera and Francesco Grossetti at Bocconi University.

Team members:

• Alessandro Caruso
• Fabian Kraus
• Silvia Juzova
• Louis Lacombe
• Riccardo Tordini
• Steffen Brockmann

Project report and slide deck are available directly on the repository.

Project video is available at: https://drive.google.com/file/d/1tiUlOwZU7c_7AnjEQ8YEJnPG16UAc9lA/view?usp=sharing

The project is to build a neural network that performs logo detection (as in object detection task) based on a large dataset of images (and noise) of 17 popular brand logos. A subset of these logos and images has been chosen and two different models, Faster R-CNN and YOLOv5, have been used to predict on a test dataset.

A detailed report highlighting the complete analysis of the models presented is provided and called Report.pdf. For optimal understanding of the different analyses taken and some further insights on why they have been taken, it is strongly advised to read through the report.

Clone the repository to your local or virtual machine. In order to run the algorithms an environment with the following settings is necessary: Python>=3.6.0 is required with all requirements.txt installed as well as PyTorch>=1.8 (for linux vm use conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch ). In order to clone some of the larger files git lfs should be installed. For more information see here: https://docs.github.com/en/repositories/working-with-files/managing-large-files/installing-git-large-file-storage.

$ git clone [email protected]:deeppaco/logo_detection_project.git

To predict with the YOLOv5 algorithm on our test datasets navigate into the yolov5 folder and run the following code:

python detect.py --source final_test/images --weights runs/exp45/weights/best.pt --img 640 --save-txt --conf-thres 0.01 --iou-thres 0.2 --save-conf

To predict on the additional test dataset we created run the following command:

python detect.py --source additional_test/images --weights runs/exp45/weights/best.pt --img 640 --save-txt --conf-thres 0.01 --iou-thres 0.2 --save-conf

The resulting predictions will be saved in the folder yolov5/runs/detect/. Use the file Post_training_analysis_yolov5.ipynb in the yolov5 folder to analyze the results (path names have to be changed!).

Additional Information

• The folders yolov5/final_test, and yolov5/additional_test contain the two test datasets we used (for more information please see the report).
• The folder yolov5/runs contains the weights of the best model we used in exp45 as well as the subfolder detect where the results of the detection on our two test datasets of the best model can be found.
• In yolov5/data/hyps the data.yaml file defines the hyperparameters we used to train the model.

All the required files for FasterRCNN can be found in the two_stage_detector folders. The model and analysis consists of two self-contained colab files that can be run simply by changing the base directory name in the file. Additionally, all the pre-model analysis are done utilising the COCO format and hence presented in the same file as the model. Further instructions can be found inside the respective files.

For preprocessing analysis, training the model and obtaining the evaluations, the FasterRCNN_logo_detection.ipynb should be run.

The above jupyter notebook will generate a csv file utilised for FasterRCNN_post_analysis.ipynb. The csv, called predictions_test_resnet_new.csv can be found in the data folder and will be automaticall loaded at the beginning. This notebook contains a detailled analysis on the results.