Install Pytorch with :

pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu116

Install the requirements with:

pip install -r requirements.txt

Download weights:

mkdir weights
cd weights
gdown 1p6LyrHIAYbz5M94eBTTOcduWF-TzeQ1D # Logo-Detection-Yolov8
gdown 1_0iA_pmC4-IvA7IaRuiJ-QHnRheIEmBe  # ECT-SAL weights
cd ..
cd saliency_prediction
mkdir pretrained_models
cd pretrained_models
gdown 1C9TgPQvFQHJP3byDtBL5X56cXeOl4JE2  # Resnet 50 pretrained weight
cd ..
cd ..

This module focuses on detecting brand logos in images using the YOLOv8 model. It utilizes two datasets for training: FoodLogoDet-1500 and LogoDet-3K.

You can use the following command to run the brand logo detection code:

python main_detection_yolov8.py --model="weights/Logo_Detection_Yolov8.pt" --image="test_images/test.jpg" --save-result

• If you want to visualize the detection results, include the --save-result flag in the command.

Original Image	Brand Logo Detection Result

This module is designed for predicting saliency maps of images, particularly suited for use in ads and packaging. Model leverages the ECSAL dataset for training. You can find the dataset here.

• For saliency map prediction, it is essential to provide a corresponding text map. Use the DBNET++ model available here to generate accurate text maps for improved saliency predictions.

Run the script:

python main_saliency_prediction.py --img_path path/to/your/image.jpg --weight_path "weights/ECT_SAL.pth" --tmap path/to/test_text_map_image.jpg --output_path path/to/output/directory

To train your dataset on the ECT-SAL model, follow the instructions provided in the ECT-SAL README.

Original Image	Saliency Map

The Brand Attention Module is a component designed to assess the visibility and attention of a brand within advertisement and packaging images. It combines logo detection and saliency map prediction techniques to quantify the presence and prominence of a brand in a given image.

• For saliency map prediction, it is essential to provide a corresponding text map. Use the DBNET++ model available here to generate accurate text maps for improved saliency predictions.

python main_brand_attention.py --img_path path/to/input_image.jpg --tmap path/to/text_map.jpg

• If the detection is accurate and aligns with the brand logo, press to confirm.

• In cases where adjustments are needed or discrepancies exist, press to signify the need for refinement.

• For refinement, use the interactive feature to draw new bounding boxes (bboxes) directly on the image. This hands-on approach allows for precise customization of logo localization.

Input Image
Brand-Attention Score: 23.54

This Module is a component designed to assess the visibility and attention of any object you want within advertisement and packaging images. It saliency map prediction techniques to quantify the presence and prominence of that object in a given image.

• For saliency map prediction, it is essential to provide a corresponding text map. Use the DBNET++ model available here to generate accurate text maps for improved saliency predictions.

python main_object_attention.py --img_path path/to/input_image.jpg --tmap path/to/text_map.jpg

• draw bounding boxes around objects of interest, press the Enter key to calculate and obtain the attention score for the selected objects.

Input Image	BBox Selected
Object Attention Score	11.22%

We thank the authors of Transalnet, DBNET, Efficient-Attention for their code repositories.

@article{hosseini2024brand,
  title={Brand Visibility in Packaging: A Deep Learning Approach for Logo Detection, Saliency-Map Prediction, and Logo Placement Analysis},
  author={Hosseini, Alireza and Hooshanfar, Kiana and Omrani, Pouria and Toosi, Reza and Toosi, Ramin and Ebrahimian, Zahra and Akhaee, Mohammad Ali},
  journal={arXiv preprint arXiv:2403.02336},
  year={2024}
}