flagsense 0.1.3

A simple package for running flag detection models on images

FlagSense

FlagSense is a Python package for detecting and classifying flags in images. Using advanced computer vision techniques, FlagSense identifies national flags and determines which country they belong to. FlagSense uses state-of-the-art models to detect flags in real-world contexts and accurately classify them by country.

Installation

FlagSense can be easily installed via pip: pip install flagsense

Usage

The following image files are accepted by YOLO and thus can be used with FlagSense: .jpg, .jpeg, .png, .bmp, .tif, .tiff, .webp.

After installation, FlagSense can be easily used in the command line by running the following code, where input_path is a filepath to an image or a folder containing many images on a local machine.

flagsense input_path

The --model flag allows the user to choose from a list of models included in the package. The default model (if no value is input) is a custom YOLOv8 model trained on all nation flags. The user can choose additional models if they prefer, or can choose continent-specific models. Continent-specific models are useful in situations where the user may know the location of an image and thus expect certain nation flags, or is only interested in classifying flags of a certain continent. For example, if a user knows they only need to search for African flags, they can use FlagSense as below by calling any Africa model they choose. The full list of models and their names are below.

flagsense input_path --model v8_africa

FlagSense can also be used within Python by importing the Detector class and calling the "detect" method on an image or folder of images.

from flagsense.yolo_inference import Detector

model = Detector("v9")
model.detect("input_path")

A list of models and the flags to call them is below:

• v8 - YOLOv8, all countries
• v9 - YOLOv9, all countries
• v10 - YOLOv10, all countries
• v8_africa: YOLOv8, Africa
• v9_africa: YOLOv9, Africa
• v10_africa: YOLOv10, Africa
• v8_asia: YOLOv8, Asia
• v9_asia: YOLOv9, Asia
• v10_asia: ,
• v8_europe: ,
• v9_europe: YOLOv9, Europe
• v10_europe: YOLOv10, Europe
• v8_northamerica: YOLOv8, North America
• v9_northamerica: YOLOv9, North America
• v10_northamerica: YOLOv10, North America
• v8_oceania: YOLOv8, Oceania
• v9_oceania: YOLOv9, Oceania
• v10_oceania: ,
• v8_southamerica: YOLOv8, South America
• v9_southamerica: YOLOv9, South America
• v10_southamerica: YOLOv10, South America

By default, FlagSense outputs annotations in both YOLO and COCO JSON format. By adding the flag “verbose”, the model can also output each image overlaid with the annotation. This is off by default since, with a large dataset, creating an additional image for each input image may increase runtime and take up storage.

flagsense input_path –verbose

Interpreting/Exporting Output (Format)

By default, FlagSense outputs annotations in YOLO and COCO JSON format. After running the package, a new folder called <input_path>_annotations will be created. This folder will contain a folder called “yolo” containing the YOLO format annotations and a folder “coco” containing the COCO JSON annotations. If the –verbose flag is used, another folder “image_overlay” will be created containing the image or images with the annotations and corresponding confidence level overlaid, as shown below.

Supported Countries and Flags

The model was trained on a dataset of 24,278 images encompassing 241 flags belonging to nations and certain territories. A list of flags recognized in the package can be found in “Supported Countries and Flags.” This includes all 193 member states recognized by the UN as well as 48 non-member states (e.g., Vatican City, Palestine) , partially recognized states (e.g., Kosovo, South Ossetia), or certain non-sovereign territories with strong national identities and recognizable flags (e.g., Puerto Rico, Greenland).

The full list of supported countries can be found in Supported Countries List.

Description of Training

Models were trained using the Ultralytics package in Python. In addition to models containing all 241 flags, we also developed continent-specific models. These models are trained on the flags of only one continent which yielded better performance. If a user knows their images are from one continent or region, using these models will yield greater performance and a smaller model with faster runtime.

All-Country Flag Detection Results

Metric	YOLOv8	YOLOv9	YOLOv10	YOLOv11	Roboflow 3.0
Precision	0.772	0.745	0.745	0.758	0.805
Recall	0.802	0.774	0.725	0.810	0.788
mAP50	0.841	0.825	0.769	0.841	0.855
mAP50-95	0.657	0.649	0.609	0.641	0.669

YOLOv8 Continent-Specific Model Results

	North America	South America	Africa	Europe	Asia	Oceania
Precision	0.810	0.896	0.884	0.762	0.854	0.888
Recall	0.834	0.825	0.907	0.676	0.803	0.840
mAP50	0.862	0.911	0.946	0.749	0.853	0.875
mAP50-95	0.639	0.653	0.748	0.540	0.631	0.730

YOLOv9 Continent-Specific Model Results

	North America	South America	Africa	Europe	Asia	Oceania
Precision	0.881	0.818	0.854	0.725	0.804	0.867
Recall	0.794	0.857	0.913	0.699	0.815	0.833
mAP50	0.862	0.894	0.959	0.770	0.850	0.883
mAP50-95	0.659	0.647	0.769	0.557	0.628	0.736

YOLOv10 Continent-Specific Model Results

	North America	South America	Africa	Europe	Asia	Oceania
Precision	0.875	0.848	0.866	0.727	0.822	0.880
Recall	0.748	0.781	0.831	0.691	0.762	0.818
mAP50	0.855	0.866	0.921	0.746	0.840	0.867
mAP50-95	0.646	0.630	0.738	0.544	0.619	0.717

Model Information and Literature Review

National flags and related symbols play a significant role in shaping political attitudes and behaviors. They can elicit emotional responses and be revealing of political beliefs (Amavilah, 2008). For example, Trofimchuk & Goh (2023) found that left-leaning individuals perceive their nation’s flags less positively than right-leaning counterparts. Becker et al. (2017) found that the meanings and attitudes associated with flags vary greatly between countries, and over time based on historical context and attitudes. They serve as political symbols that represent nationality and can be both uniting and divisive in their usage (Elgenius, 2011). As such, the detection of national flags using computer vision can glean political insights into environments and locations.

Recent research has explored the application of convolutional neural networks (CNNs) for national flag detection and recognition tasks in computer vision. Flags present a unique challenge for object detection because they often appear in dynamic settings—waving on poles, displayed on clothing, or in complex image scenes—and many national flags have subtle design differences (Gu et al., 2018). CNNs, which have revolutionized object detection in computer vision, surpass traditional methods by learning hierarchical features directly from data rather than relying on handcrafted features (Neha et al., 2024; Zhiqiang Wang & Liu Jun, 2017). These capabilities make CNNs particularly well-suited for detecting flags, as they can discern intricate patterns and variations in flag designs. Object detection models based on CNNs can be categorized into two approaches: two-stage approaches like R-CNN and Faster R-CNN and one-stage, like YOLO and RetinaNET (F. Sultana et al., 2019). These models have exhibited strong performances on benchmark datasets like PASCAL VOC and MS COCO (Kaur & Singh, 2022).

Kulik & Shtanko (2020) investigated CNNs for recognizing complex objects with varied appearances by focusing on national flags, noting their diverse representations in images. Yerbolat et al. (2024) conducted a comparative analysis of YOLOv8 and Faster R-CNN algorithms for national flag recognition, finding that YOLOv8 offers high processing speed but lower accuracy at high IoU, while Faster R-CNN provides better accuracy at high IoU but slower processing. Gu et al. (2018) developed a CNN using a robust dataset containing flags of different sizes, occlusions, lighting conditions, and deformations and noted greater detection performance compared to other models. However, they noted the potential for improvement using a larger dataset, as the study used a dataset containing only 1668 positive samples. Further studies could benefit from the testing of additional models, like RetinaNET, EfficientDET, and newer YOLO models, as comparisons between models are generally heavily dependent on the specific task (Sanchez et al., 2020). Cabrera-Ponce & Martínez-Carranza (2020) implemented an onboard CNN-based system for flag detection in autonomous landing tasks for micro aerial vehicles (MAVs), achieving high confidence metrics for both flag and landing platform detection. Their approach outperformed conventional computer vision methods like template and feature matching. While this CNN does not involve the detection of national flags, it demonstrates the effectiveness of CNNs in addressing challenges such as illumination changes, rotation, and scale variations in flag detection scenarios.

Despite the existence of convolutional neural networks for flag detection, there remains a gap in the availability of an easy-to-use Python package that provides fast and accessible national flag detection in images. Flagnet is a publicly available neural network capable of detecting flags for 193 United Nations member countries. National Flag Recognition using Machine Learning Techniques is a similar project consisting of 290 countries. Likewise, some Python packages concerning national flags exist, but none can identify and classify flags from images. Pycountry and django-countries are data lookup tools that can retrieve country data, including flags, but no packages exist that can detect flags from images using a complex CNN. While projects like Flagnet and similar deep learning implementations exist for flag detection, they are primarily standalone models or prototypes that require manual setup and adaptation. A dedicated Python package would offer a standardized, well-documented, and easy-to-install library for flag detection, enabling researchers and developers to quickly incorporate this functionality into their workflows without extensive customization. Such a package could provide both high-performance models and practical usability, fostering broader and more accessible adoption and enabling diverse use cases such as automated image analysis and educational tools.

License

DeepFace is licensed under the MIT License - see LICENSE for more details.