Object detection is a computer vision technique that allows us to identify and locate objects in an image or video. With this kind of identification and localization, object detection can be used to count objects in a scene and determine and track their precise locations while accurately labeling them. Object detection is commonly confused with image recognition, so before we proceed, it’s important that we clarify the distinctions between them. Image recognition assigns a label to an image. A picture of a dog receives the label “dog”. A picture of two dogs still receives the label “dog”. Object detection, on the other hand, draws a box around each dog and labels the box “dog”. The model predicts where each object is and what label should be applied. The purpose of this project is training a logo detection model with YOLOv7 using two different datasets. It can detect logos in the wild images.
In this model, we used YOLOv7 as the architecture. YOLOv7 surpasses all known object detectors in both speed and accuracy in the range from 5 FPS to 160 FPS and has the highest accuracy 56.8% AP among all known real-time object detectors with 30 FPS or higher on GPU V100. YOLO architecture is FCNN(Fully Connected Neural Network) based. However, Transformer-based versions have also recently been added to the YOLO family.
The YOLO has three main components.
- Backbone
- Head
- Neck
The Backbone mainly extracts essential features of an image and feeds them to the Head through Neck. The Neck collects feature maps extracted by the Backbone and creates feature pyramids. Finally, the head consists of output layers that have final detections.
YOLOv7 improves speed and accuracy by introducing several architectural reforms. The following major changes have been introduced in the YOLOv7 paper.
-
Architectural Reforms
- E-ELAN (Extended Efficient Layer Aggregation Network)
- Model Scaling for Concatenation-based Models
-
Trainable BoF (Bag of Freebies)
- Planned re-parameterized convolution
- Coarse for auxiliary and Fine for lead loss
You can see detailed information about these additions in this paper.
We have two datasets in this work. One is smaller than the other. The small one is called Flickr Logos 27. he training set contains 810 annotated images, corresponding to 27 logo classes/brands (30 images for each class). All images are annotated with bounding boxes of the logo instances in the image.
The brands included in the dataset are Adidas, Apple, BMW, Citroen, Coca Cola, DHL, Fedex, Ferrari, Ford, Google, Heineken, HP, McDonalds, Mini, Nbc, Nike, Pepsi, Porsche, Puma, Red Bull, Sprite, Starbucks, Intel, Texaco, Unisef, Vodafone and Yahoo.
To see the details of the Flickr Logos 27 dataset, please visit this page.
The other dataset is LogoDet-3K. LogoDet-3K is the largest logo detection dataset with full annotation, which has 3,000 logo categories, about 200,000 manually annotated logo objects, and 158,652 images. To see which brands are included, please visit this page.
The model is trained using Flickr Logos 27 because LogoDet-3K training would take too much time. However, this project is ready to be trained with LogoDet-3K. The steps to use LogoDet-3K will be added soon.
Instructions on setting up your project locally. To get a local copy up and running follow these simple steps.
To download the dataset, run getFlickr.sh.
sh getFlickr.shIt will be downloaded inside to data folder.
In this project, three are three base models used. You can download them by visiting the links below.
Install submodules
git submodule update --initTo install the required packages. In a terminal, type:
pip install -r src/requirements.txtNow that we have our dataset, we need to convert the annotations into the format expected by YOLOv7. YOLOv7 expects data to be organized in a specific way, otherwise it is unable to parse through the directories.
python src/convert_annotations.py --dataset flickr27To see if the conversion is correct, run.
python src/convert_annotations.py --dataset flickr27 --plotNext, we need to partition the dataset into train, validation, and test sets. These will contain 80%, 10%, and 10% of the data, respectively.
python src/prepare_data.py --dataset flickr27The training specifications are:
- Epoch: 300
- Dataset: Flickr Logos 27
- Batch size: 2
- Image size: 640
- GPU: NVIDIA GeForce RTX 3060 Laptop GPU
If you are having fitting the model into the memory:
- Use a smaller batch size.
- Use a smaller network: the yolov7-tiny.pt checkpoint will run at lower cost than the basic yolov7_training.pt.
- Use a smaller image size: the size of the image corresponds directly to expense during training. Reduce the images from 640 to 320 to significantly cut cost at the expense of losing prediction accuracy.
To start the training:
python src/yolov7/train.py --img-size 640 --cfg src/cfg/training/yolov7.yaml --hyp data/hyp.scratch.yaml --batch 2 --epoch 300 --data data/logo_data_flickr.yaml --weights src/yolov7_training.pt --workers 2 --name yolo_logo_det --device 0You can also train the model on Google Colab.
To test the training model:
python src/yolov7/detect.py --source data/Sample/test --weights runs/train/yolo_logo_det/weights/best.pt --conf 0.25 --name yolo_logo_detDownload the resulting model here.
Confusion Matrix:
PR Curve:
