A Dataset comprised of two parts, images generated by AI image generation models such as DALL-E and Midjourney, and real images known to be made by humans. The majority of AI generated images are artistic works of some type and not photorealistic because it was found that having more artistic works than photos in the human generated set yielded better test results. One major issue found when trying to train classifiers on this set is while a test accuracy as high as 94% was achieved, if the image (regardless of source AI or human) contained noise such as a film grain or fur there was a higher error rate and the image was more likely to be mislabeled as AI generated. My theory is because diffusion image generation models (DALL-E etc.) start with random noise and turn it into an image based on the prompt, so the classifier could be using the noise of the image as a way to detect Ai generated art and by adding noise the model is getting confused. One possible solution to this is using image denoising on the image or edge detection however I have yet to test either.

Benefits over other datasets

The benefit of this dataset compared to other artificially generated image datasets (such as CIFAKE) is that all images are in there original size and aspect ratio.

CIFAKE: Real and AI-Generated Synthetic Images

The quality of AI-generated images has rapidly increased, leading to concerns of authenticity and trustworthiness.

CIFAKE is a dataset that contains 60,000 synthetically-generated images and 60,000 real images (collected from CIFAR-10). Can computer vision techniques be used to detect when an image is real or has been generated by AI?

Further information on this dataset can be found here: Bird, J.J. and Lotfi, A., 2024. CIFAKE: Image Classification and Explainable Identification of AI-Generated Synthetic Images. IEEE Access.

Dataset details

The dataset contains two classes - REAL and FAKE.

For REAL, we collected the images from Krizhevsky & Hinton's CIFAR-10 dataset

For the FAKE images, we generated the equivalent of CIFAR-10 with Stable Diffusion version 1.4

There are 100,000 images for training (50k per class) and 20,000 for testing (10k per class)

Papers with Code

The dataset and all studies using it are linked using Papers with Code https://paperswithcode.com/dataset/cifake-real-and-ai-generated-synthetic-images

References

If you use this dataset, you must cite the following sources

Krizhevsky, A., & Hinton, G. (2009). Learning multiple layers of features from tiny images.

Bird, J.J. and Lotfi, A., 2024. CIFAKE: Image Classification and Explainable Identification of AI-Generated Synthetic Images. IEEE Access.

Real images are from Krizhevsky & Hinton (2009), fake images are from Bird & Lotfi (2024). The Bird & Lotfi study is available here.

Notes

The updates to the dataset on the 28th of March 2023 did not change anything; the file formats ".jpeg" were renamed ".jpg" and the root folder was uploaded to meet Kaggle's usability requirements.

License

This dataset is published under the same MIT license as CIFAR-10:

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

The dataset is a captivating ensemble of images sourced from two distinct channels: web scraping and AI-generated content. The content covers many subjects; however, special emphasis was placed on these topics: people, animals, portraits, scenery, and psychedelics.

Key Features:

Web-Scraped Images: These images are harvested from various online sources across the web. Ranging from landscapes, paintings, psychedelic trips, and portraits, the web-scraped images offer a glimpse into the vast spectrum of digital imagery available online.

Projects and Applications:

Image Classification and Recognition: Researchers and developers can leverage the dataset to train machine learning models for image classification and recognition tasks. By incorporating both web-scraped and AI-generated images, models can learn to identify and categorize objects, scenes, and concepts across diverse domains with greater accuracy and generalization.

Artistic Exploration and Creative Synthesis: Artists, designers, and creative enthusiasts can draw inspiration from the dataset to explore new avenues of artistic expression and experimentation. They can use AI-generated imagery as a canvas for artistic reinterpretation, blending traditional techniques with computational aesthetics to produce captivating artworks and multimedia installations.

Data Visualization and Exploratory Analysis: Data scientists and researchers can analyze the dataset to uncover insights into visual trends, patterns, and correlations.

Have fun!

Dataset described in the paper "Synthbuster: Towards Detection of Diffusion Model Generated Images" (Quentin Bammey, 2023, Open Journal of Signal Processing)

This dataset contains synthetic, AI-generated images from 9 different models:

• DALL·E 2
• DALL·E 3
• Adobe Firefly
• Midjourney v5
• Stable Diffusion 1.3
• Stable Diffusion 1.4
• Stable Diffusion 2
• Stable Diffusion XL
• Glide

1000 images were generated per model. The images are loosely based on raise-1k images (Dang-Nguyen, Duc-Tien, et al. "Raise: A raw images dataset for digital image forensics." Proceedings of the 6th ACM multimedia systems conference. 2015.). For each image of the raise-1k dataset, a description was generated using the Midjourney /describe function and CLIP interrogator (https://github.com/pharmapsychotic/clip-interrogator/). Each of these prompts was manually edited to produce results as photorealistic as possible and remove living persons and artists names.

In addition to this, parameters were randomly selected within reasonable values for methods requiring so.

The prompts and parameters used for each method can be found in the `prompts.csv` file.

This dataset can be used to evaluate AI-generated image detection methods. We recommend matching the generated images with the real Raise-1k images, to evaluate whether the methods can distinguish the two of them. Raise-1k images are not included in the dataset, they can be downloaded separately at (http://loki.disi.unitn.it/RAISE/download.html).

None of the images suffered degradations such as JPEG compression or resampling, which leaves room to add your own degradations to test robustness to various transformation in a controlled manner.

Data 1 AI Generated Images

## Overview

Data 1  AI Generated Images is a dataset for object detection tasks - it contains Houses annotations for 258 images.

## Getting Started

You can download this dataset for use within your own projects, or fork it into a workspace on Roboflow to create your own model.

  ## License

  This dataset is available under the [CC BY 4.0 license](https://creativecommons.org/licenses/CC BY 4.0).

A dataset combining AI-generated and web-scraped images across subjects like people, animals, portraits, scenery, and psychedelics. Designed for image classification, recognition, and creative AI research.

This dataset features over 80,000 high-quality images of construction sites sourced from photographers worldwide. Built to support AI and machine learning applications, it delivers richly annotated and visually diverse imagery capturing real-world construction environments, machinery, and processes.

Key Features: 1. Comprehensive Metadata: the dataset includes full EXIF data such as aperture, ISO, shutter speed, and focal length. Each image is annotated with construction phase, equipment types, safety indicators, and human activity context—making it ideal for object detection, site monitoring, and workflow analysis. Popularity metrics based on performance on our proprietary platform are also included.

1. Unique Sourcing Capabilities: images are collected through a proprietary gamified platform, with competitions focused on industrial, construction, and labor themes. Custom datasets can be generated within 72 hours to target specific scenarios, such as building types, stages (excavation, framing, finishing), regions, or safety compliance visuals.

2. Global Diversity: sourced from contributors in over 100 countries, the dataset reflects a wide range of construction practices, materials, climates, and regulatory environments. It includes residential, commercial, industrial, and infrastructure projects from both urban and rural areas.

3. High-Quality Imagery: includes a mix of wide-angle site overviews, close-ups of tools and equipment, drone shots, and candid human activity. Resolution varies from standard to ultra-high-definition, supporting both macro and contextual analysis.

4. Popularity Scores: each image is assigned a popularity score based on its performance in GuruShots competitions. These scores provide insight into visual clarity, engagement value, and human interest—useful for safety-focused or user-facing AI models.

5. AI-Ready Design: this dataset is structured for training models in real-time object detection (e.g., helmets, machinery), construction progress tracking, material identification, and safety compliance. It’s compatible with standard ML frameworks used in construction tech.

6. Licensing & Compliance: fully compliant with privacy, labor, and workplace imagery regulations. Licensing is transparent and ready for commercial or research deployment.

Use Cases: 1. Training AI for safety compliance monitoring and PPE detection. 2. Powering progress tracking and material usage analysis tools. 3. Supporting site mapping, autonomous machinery, and smart construction platforms. 4. Enhancing augmented reality overlays and digital twin models for construction planning.

This dataset provides a comprehensive, real-world foundation for AI innovation in construction technology, safety, and operational efficiency. Custom datasets are available on request. Contact us to learn more!

ShutterStock AI vs. Human-Generated Image Dataset

This dataset is curated to facilitate research in distinguishing AI-generated images from human-created ones, leveraging ShutterStock data. As AI-generated imagery becomes more sophisticated, developing models that can classify and analyze such images is crucial for applications in content moderation, digital forensics, and media authenticity verification.

Dataset Overview:

• Total Images: 100,000
• Training Data: 80,000 images (majority AI-generated)
• Test Data: 20,000 images
• Image Sources: A mix of AI-generated images and real photographs or illustrations created by human artists
• Labeling: Each image is labeled as either AI-generated or human-created

Potential Use Cases:

• AI-Generated Image Detection: Train models to distinguish between AI and human-made images.
• Deep Learning & Computer Vision Research: Develop and benchmark CNNs, transformers, and other architectures.
• Generative Model Evaluation: Compare AI-generated images to real images for quality assessment.
• Digital Forensics: Identify synthetic media for applications in fake image detection.
• Ethical AI & Content Authenticity: Study the impact of AI-generated visuals in media and ensure transparency.

Why This Dataset?

With the rise of generative AI models like Stable Diffusion, DALL·E, and MidJourney, the ability to differentiate between synthetic and real images has become a crucial challenge. This dataset offers a structured way to train AI models on this task, making it a valuable resource for both academic research and practical applications.

Explore the dataset and contribute to advancing AI-generated content detection!

Step 1: Install and Authenticate Kaggle API

If you haven't installed the Kaggle API, run:
bash pip install kaggle Then, download your kaggle.json API key from Kaggle Account and move it to ~/.kaggle/ (Linux/Mac) or `C:\Users\YourUser.kaggle` (Windows).

Step 2: Use wget

  wget --no-check-certificate --header "Authorization: Bearer $(cat ~/.kaggle/kaggle.json | jq -r .token)" "https://www.kaggle.com/datasets/shreyasraghav/shutterstock-dataset-for-ai-vs-human-gen-image" -O dataset.zip

Step 3: Extract the Dataset

Once downloaded, extract the dataset using:
bash unzip dataset.zip -d dataset_folder

Now your dataset is ready to use! 🚀

Diffusion models (DMs) have revolutionized image generation, producing high-quality images with applications spanning various fields. However, their ability to create hyper-realistic images poses significant challenges in distinguishing between real and synthetic content, raising concerns about digital authenticity and potential misuse in creating deepfakes. This work introduces a robust detection framework that integrates image and text features extracted by CLIP model with a Multilayer Perceptron (MLP) classifier.

The dataset for this project is represented by photos, photos for the buildings of the University of Salford, these photos are taken by a mobile phone camera from different angels and different distances , even though this task sounds so easy but it encountered some challenges, these challenges are summarized below: 1. Obstacles. a. Fixed or unremovable objects. When taking several photos for a building or a landscape from different angels and directions ,there are some of these angels blocked by a form of a fixed object such as trees and plants, light poles, signs, statues, cabins, bicycle shades, scooter stands, generators/transformers, construction barriers, construction equipment and any other service equipment so it is unavoidable to represent some photos without these objects included, this will raise 3 questions. - will these objects confuse the model/application we intend to create meaning will that obstacle prevent the model/application from identifying the designated building? - Or will the photos be more precise with these objects and provide the capability for the model/application to identify these building with these obstacles included? - How far is the maximum length for detection? In other words, how far will the mobile device with the application be from the building so it could or could not detect the designated building? b. Removable and moving objects. - Any University is crowded with staff and students especially in the rush hours of the day so it is hard for some photos to be taken without a personnel appearing in that photo in a certain time period of the day. But, due to privacy issues and showing respect to that person, these photos are better excluded. - Parked vehicles, trollies and service equipment can be an obstacle and might appear in these images as well as it can block access to some areas which an image from a certain angel cannot be obtained. - Animals, like dogs, cats, birds or even squirrels cannot be avoided in some photos which are entitled to the same questions above.
2. Weather. In a deep learning project, more data means more accuracy and less error, at this stage of our project it was agreed to have 50 photos per building but we can increase the number of photos for more accurate results but due to the limitation of time for this project it was agreed for 50 per building only. these photos were taken on cloudy days and to expand our work on this project (as future works and recommendations). Photos on sunny, rainy, foggy, snowy and any other weather condition days can be included. Even photos in different times of the day can be included such as night, dawn, and sunset times. To provide our designated model with all the possibilities to identify these buildings in all available circumstances.

1. The selected buildings. It was agreed to select 10 buildings only from the University of Salford buildings for this project with at least 50 images per building, these selected building for this project with the number of images taken are:
2. Chapman: 74 images
3. Clifford Whitworth Library: 60 images
4. Cockcroft: 67 images
5. Maxwell: 80 images
6. Media City Campus: 92 images
7. New Adelphi: 93 images
8. New Science, Engineering & Environment: 78 images
9. Newton: 92 images
10. Sports Centre: 55 images

11. University House: 60 images Peel building is an important figure of the University of Salford due to its distinct and amazing exterior design but unfortunately it was excluded from the selection due to some maintenance activities at the time of collecting the photos for this project as it is partially covered with scaffolding and a lot of movement by personnel and equipment. If the supervisor suggests that this will be another challenge to include in the project then, it is mandatory to collect its photos. There are many other buildings in the University of Salford and again to expand our project in the future, we can include all the buildings of the University of Salford. The full list of buildings of the university can be reviewed by accessing an interactive map on: www.salford.ac.uk/find-us

12. Expand Further. This project can be improved furthermore with so many capabilities, again due to the limitation of time given to this project , these improvements can be implemented later as future works. In simple words, this project is to create an application that can display the building’s name when pointing a mobile device with a camera to that building. Future featured to be added: a. Address/ location: this will require collection of additional data which is the longitude and latitude of each building included or the post code which will be the same taking under consideration how close these buildings appear on the interactive map application such as Google maps, Google earth or iMaps. b. Description of the building: what is the building for, by which school is this building occupied? and what facilities are included in this building? c. Interior Images: all the photos at this stage were taken for the exterior of the buildings, will interior photos make an impact on the model/application for example, if the user is inside newton or chapman and opens the application, will the building be identified especially the interior of these buildings have a high level of similarity for the corridors, rooms, halls, and labs? Will the furniture and assets will be as obstacles or identification marks? d. Directions to a specific area/floor inside the building: if the interior images succeed with the model/application, it would be a good idea adding a search option to the model/application so it can guide the user to a specific area showing directions to that area, for example if the user is inside newton building and searches for lab 141 it will direct him to the first floor of the building with an interactive arrow that changes while the user is approaching his destination. Or, if the application can identify the building from its interior, a drop down list will be activated with each floor of this building, for example, if the model/application identifies Newton building, the drop down list will be activated and when pressing on that drop down list it will represent interactive tabs for each floor of the building, selecting one of the floors by clicking on its tab will display the facilities on that floor for example if the user presses on floor 1 tab, another screen will appear displaying which facilities are on that floor. Furthermore, if the model/application identifies another building, it should activate a different number of floors as buildings differ in the number of floors from each other. this feature can be improved with a voice assistant that can direct the user after he applies his search (something similar to the voice assistant in Google maps but applied to the interior of the university’s buildings. e. Top View: if a drone with a camera can be afforded, it can provide arial images and top views for the buildings that can be added to the model/application but these images can be similar to the interior images situation , the buildings can be similar to each other from the top with other obstacles included like water tanks and AC units.

13. Other Questions:

14. Will the model/application be reproducible? the presumed answer for this question should be YES, IF, the model/application will be fed with the proper data (images) such as images of restaurants, schools, supermarkets, hospitals, government facilities...etc.

Objective In response to the risks of highly realistic image misuse arising from the rapid development of Artificial Intelligence Generated Content (AIGC) technology, and the challenges of existing detection methods struggling to adapt to continuously emerging new generative models and lacking continual learning capabilities, this paper constructs the first benchmark dataset for the continual detection of AI-generated images to address this challenge and proposes a corresponding continual detection framework.Methods First, we constructed a benchmark dataset for continual learning in AI-generated image detection, which includes samples from five mainstream generative models as well as real images, and is organized into a continual learning task stream. Second, we systematically defined and investigated the challenges faced by continual learning in this detection task, with a special focus on a novel "mixed binary- and single-class" incremental learning scenario that reflects real-world constraints. Based on this, we established three benchmarks with varying degrees of sample replay constraints. Finally, we adapted existing continual learning methods for each benchmark scenario and proposed a universal conversion framework for the most stringent no-replay setting to restore the efficacy of methods that fail under this condition.Results Experiments conducted on our proposed dataset validate the effectiveness of the benchmark and the methods. In scenarios permitting replay, the adapted methods successfully achieve incremental detection. In the strictest no-replay scenario, traditional non-replay methods suffer from severe performance degradation or even fail completely. In contrast, the application of our proposed universal conversion framework leads to a significant performance boost for these methods, effectively enhancing detection accuracy and source identification capabilities while substantially mitigating catastrophic forgetting.Conclusion This paper successfully constructs a benchmark for the continual detection of AI-generated images, provides an in-depth analysis of the key challenges involved, and proposes effective continual detection strategies and solutions, notably introducing an innovative framework for continual learning in no-replay scenarios. The findings of this research offer crucial methodological support and empirical evidence for the development of robust and adaptive detection systems capable of keeping pace with the ever-evolving landscape of AI generation technologies.

Round 13 Train DatasetThis is the training data used to create and evaluate trojan detection software solutions. This data, generated at NIST, consists of object detection AIs trained both on synthetic image data build from Cityscapes and the DOTA_v2 dataset. A known percentage of these trained AI models have been poisoned with a known trigger which induces incorrect behavior. This data will be used to develop software solutions for detecting which trained AI models have been poisoned via embedded triggers. This dataset consists of 128 AI models using a small set of model architectures. Half (50%) of the models have been poisoned with an embedded trigger which causes misclassification of the input when the trigger is present.

GENERATED USA Passports Segmentation

The dataset contains a collection of images representing GENERATED USA Passports. Each passport image is segmented into different zones, including the passport zone, photo, name, surname, date of birth, sex, nationality, passport number, and MRZ (Machine Readable Zone). The dataset can be utilized for computer vision, object detection, data extraction and machine learning models. Generated passports can assist in conducting research without… See the full description on the dataset page: https://huggingface.co/datasets/UniqueData/generated-passports-segmentation.

Object Detection Data Labeling

## Overview

Object Detection Data Labeling is a dataset for object detection tasks - it contains Objects annotations for 285 images.

## Getting Started

You can download this dataset for use within your own projects, or fork it into a workspace on Roboflow to create your own model.

  ## License

  This dataset is available under the [CC BY 4.0 license](https://creativecommons.org/licenses/CC BY 4.0).

Round 11 Train DatasetThis is the training data used to create and evaluate trojan detection software solutions. This data, generated at NIST, consists of image classification AIs trained on synthetic image data build from Cityscapes. A known percentage of these trained AI models have been poisoned with a known trigger which induces incorrect behavior. This data will be used to develop software solutions for detecting which trained AI models have been poisoned via embedded triggers. This dataset consists of 288 AI models using a small set of model architectures. Half (50%) of the models have been poisoned with an embedded trigger which causes misclassification of the input when the trigger is present.

Semantic segmentation results using only one type of the real underwater sonar image datasets and synthetic underwater sonar image datasets.

Limited training data is one of the biggest challenges in the industrial application of deep learning. Generating synthetic training images is a promising solution in computer vision; however, minimizing the domain gap between synthetic and real-world images remains a problem. Therefore, based on a real-world application, we explored the generation of images with physics-based rendering for an industrial object detection task. Setting up the render engine’s environment requires a lot of choices and parameters. One fundamental question is whether to apply the concept of domain randomization or use domain knowledge to try and achieve photorealism. To answer this question, we compared different strategies for setting up lighting, background, object texture, additional foreground objects and bounding box computation in a data-centric approach. We compared the resulting average precision from generated images with different levels of realism and variability. In conclusion, we found that domain randomization is a viable strategy for the detection of industrial objects. However, domain knowledge can be used for object-related aspects to improve detection performance. Based on our results, we provide guidelines and an open-source tool for the generation of synthetic images for new industrial applications.

Introducing the English Product Image Dataset - a diverse and comprehensive collection of images meticulously curated to propel the advancement of text recognition and optical character recognition (OCR) models designed specifically for the English language.

Dataset Contain & Diversity

Containing a total of 2000 images, this English OCR dataset offers diverse distribution across different types of front images of Products. In this dataset, you'll find a variety of text that includes product names, taglines, logos, company names, addresses, product content, etc. Images in this dataset showcase distinct fonts, writing formats, colors, designs, and layouts.

To ensure the diversity of the dataset and to build a robust text recognition model we allow limited (less than five) unique images from a single resource. Stringent measures have been taken to exclude any personally identifiable information (PII) and to ensure that in each image a minimum of 80% of space contains visible English text.

Images have been captured under varying lighting conditions – both day and night – along with different capture angles and backgrounds, to build a balanced OCR dataset. The collection features images in portrait and landscape modes.

All these images were captured by native English people to ensure the text quality, avoid toxic content and PII text. We used the latest iOS and Android mobile devices above 5MP cameras to click all these images to maintain the image quality. In this training dataset images are available in both JPEG and HEIC formats.

Metadata

Along with the image data, you will also receive detailed structured metadata in CSV format. For each image, it includes metadata like image orientation, county, language, and device information. Each image is properly renamed corresponding to the metadata.

The metadata serves as a valuable tool for understanding and characterizing the data, facilitating informed decision-making in the development of English text recognition models.

Update & Custom Collection

We're committed to expanding this dataset by continuously adding more images with the assistance of our native English crowd community.

If you require a custom product image OCR dataset tailored to your guidelines or specific device distribution, feel free to contact us. We're equipped to curate specialized data to meet your unique needs.

Furthermore, we can annotate or label the images with bounding box or transcribe the text in the image to align with your specific project requirements using our crowd community.

License

This Image dataset, created by FutureBeeAI, is now available for commercial use.

Conclusion:

Leverage the power of this product image OCR dataset to elevate the training and performance of text recognition, text detection, and optical character recognition models within the realm of the English language. Your journey to enhanced language understanding and processing starts here.

The INDIGO Change Detection Reference Dataset

Description

This graffiti-centred change detection dataset was developed in the context of INDIGO, a research project focusing on the documentation, analysis and dissemination of graffiti along Vienna's Donaukanal. The dataset aims to support the development and assessment of change detection algorithms.

The dataset was collected from a test site approximately 50 meters in length along Vienna's Donaukanal during 11 days between 2022/10/21 and 2022/12/01. Various cameras with different settings were used, resulting in a total of 29 data collection sessions or "epochs" (see "EpochIDs.jpg" for details). Each epoch contains 17 images generated from 29 distinct 3D models with different textures. In total, the dataset comprises 6,902 unique image pairs, along with corresponding reference change maps. Additionally, exclusion masks are provided to ignore parts of the scene that might be irrelevant, such as the background.

To summarise, the dataset, labelled as "Data.zip," includes the following:

• Synthetic Images: These are colour images created within Agisoft Metashape Professional 1.8.4, generated by rendering views from 17 artificial cameras observing 29 differently textured versions of the same 3D surface model.
• Change Maps: Binary images that were manually and programmatically generated, using a Python script, from two synthetic graffiti images. These maps highlight the areas where changes have occurred.
• Exclusion Masks: Binary images are manually created from synthetic graffiti images to identify "no data" areas or irrelevant ground pixels.

Image Acquisition

Image acquisition involved the use of two different camera setups. The first two datasets (ID 1 and 2; cf. "EpochIDs.jpg") were obtained using a Nikon Z 7II camera with a pixel count of 45.4 MP, paired with a Nikon NIKKOR Z 20 mm lens. For the remaining image datasets (ID 3-29), a triple GoPro setup was employed. This triple setup featured three GoPro cameras, comprising two GoPro HERO 10 cameras and one GoPro HERO 11, all securely mounted within a frame. This triple-camera setup was utilised on nine different days with varying camera settings, resulting in the acquisition of 27 image datasets in total (nine days with three datasets each).

Data Structure

The "Data.zip" file contains two subfolders:

• 1_ImagesAndChangeMaps: This folder contains the primary dataset. Each subfolder corresponds to a specific epoch. Within each epoch folder resides a subfolder for every other epoch with which a distinct epoch pair can be created. It is important to note that the pairs "Epoch Y and Epoch Z" are equivalent to "Epoch Z and Epoch Y", so the latter combinations are not included in this dataset. Each sub-subfolder, organised by epoch, contains 17 more subfolders, which hold the image data. These subfolders consist of:
  • Two synthetic images rendered from the same synthetic camera ("X_Y.jpg" and "X_Z.jpg")
  • The corresponding binary reference change map depicting the graffiti-related differences between the two images ("X_YZ.png"). Black areas denote new graffiti (i.e. "change"), and white denotes "no change". "DataStructure.png" provides a visual explanation concerning the creation of the dataset.
    
    The filenames follow the following pattern:
    • X - Is the ID number of the synthetic camera. In total, 17 synthetic cameras were placed along the test site
    • Y - Corresponds to the reference epoch (i.e. the "older epoch")
    • Z - Corresponds to the "new epoch"
      
• 2_ExclusionMasks: This folder contains the binary exclusion masks. They were manually created from synthetic graffiti images and identify "no data" areas or areas considered irrelevant, such as "ground pixels". Two exclusion masks were generated for each of the 17 synthetic cameras:
  • "groundMasks": depict ground pixels which are usually irrelevant for the detection of graffiti
  • "noDataMasks": depict "background" for which no data is available.
    

A detailed dataset description (including detailed explanations of the data creation) is part of a journal paper currently in preparation. The paper will be linked here for further clarification as soon as it is available.

Licensing

Due to the nature of the three image types, this dataset comes with two licenses:

• Synthetic images:
  • These come with an In Copyright license (for the rights usage terms, see https://rightsstatements.org/page/InC/1.0/?language=en).
  • The copyright lies with:
    • the Ludwig Boltzmann Gesellschaft (https://d-nb.info/gnd/1024204324)
    • the TU Wien (https://d-nb.info/gnd/55426-1)
    • One or more anonymous graffiti creator(s) upon whose work these images are based.
  • The first two entities are also the licensor of these images.
    
• Change maps and masks:
  • These are openly licensed via CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0)
  • In this case, the copyright lies with:
    • the Ludwig Boltzmann Gesellschaft (https://d-nb.info/gnd/1024204324)
    • the TU Wien (https://d-nb.info/gnd/55426-1)
  • Both institutes are also the licensor of these images.

Every synthetic image, change map and mask has this licensing information embedded as IPTC photo metadata. In addition, the images' IPTC metadata also provide a short image description, the image creator and the creator's identity (in the form of an ORCiD).

-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

If there are any questions, problems or suggestions for the dataset or the description, please do not hesitate to contact the corresponding author, Benjamin Wild.

Introducing the Bahasa Product Image Dataset - a diverse and comprehensive collection of images meticulously curated to propel the advancement of text recognition and optical character recognition (OCR) models designed specifically for the Bahasa language.

Dataset Contain & Diversity

Containing a total of 2000 images, this Bahasa OCR dataset offers diverse distribution across different types of front images of Products. In this dataset, you'll find a variety of text that includes product names, taglines, logos, company names, addresses, product content, etc. Images in this dataset showcase distinct fonts, writing formats, colors, designs, and layouts.

To ensure the diversity of the dataset and to build a robust text recognition model we allow limited (less than five) unique images from a single resource. Stringent measures have been taken to exclude any personally identifiable information (PII) and to ensure that in each image a minimum of 80% of space contains visible Bahasa text.

Images have been captured under varying lighting conditions – both day and night – along with different capture angles and backgrounds, to build a balanced OCR dataset. The collection features images in portrait and landscape modes.

All these images were captured by native Bahasa people to ensure the text quality, avoid toxic content and PII text. We used the latest iOS and Android mobile devices above 5MP cameras to click all these images to maintain the image quality. In this training dataset images are available in both JPEG and HEIC formats.

Metadata

Along with the image data, you will also receive detailed structured metadata in CSV format. For each image, it includes metadata like image orientation, county, language, and device information. Each image is properly renamed corresponding to the metadata.

The metadata serves as a valuable tool for understanding and characterizing the data, facilitating informed decision-making in the development of Bahasa text recognition models.

Update & Custom Collection

We're committed to expanding this dataset by continuously adding more images with the assistance of our native Bahasa crowd community.

If you require a custom product image OCR dataset tailored to your guidelines or specific device distribution, feel free to contact us. We're equipped to curate specialized data to meet your unique needs.

Furthermore, we can annotate or label the images with bounding box or transcribe the text in the image to align with your specific project requirements using our crowd community.

License

This Image dataset, created by FutureBeeAI, is now available for commercial use.

Conclusion:

Leverage the power of this product image OCR dataset to elevate the training and performance of text recognition, text detection, and optical character recognition models within the realm of the Bahasa language. Your journey to enhanced language understanding and processing starts here.