Gratitude Services Case Study: Pioneering AI-Powered Cemetery Mapping in Estonia and Beyond

The problem described

The vast areas of 325 cemeteries should be well described and mapped in order to reduce the time and cost of services provided. Previously, all the mapping was done manually, requiring months of work. The goal was to create a Google Street View-like experience of a cemetery with automated OCR and object-detection AI technologies.

Allan Selirand, Gratitude Services co-founder says that the team is not limiting their aspirations to just Estonia. They envision scaling the services across borders and reaching out to different countries.

We’re leveraging the power of AI to accelerate our mapping and growth processes. Our ambitious aim is to create the world’s largest cemetery database – an initiative that requires swift, automated processes. Every single day, we’re making strides to expand our services, creating an invaluable resource for those who wish to honor and remember their departed loved ones. (Allan Selirand, Gratitude Services co-founder)

Creating the model

The cornerstone of our endeavor was to harness the power of AI models for detailed mapping and efficient provision of location-specific information about tombstones and their respective inscriptions. Creating an AI model that would seamlessly blend with our requirements meant both careful consideration and a strategic approach. The subsequent process revolved around an intricate interplay of various factors.

AI models described

First, the pictures taken in the cemeteries go to the object-detecting model, where tombstones are detected. As the AI model needed hi-res pictures for better detection accuracy, the pipeline was built to anonymize all images and only use 360-degree images from the three cameras: one front camera and two on either side. To map GPS data into the real world, we required an additional step. As the high resolution of the images slowed the labeling interface, lower-quality images were automatically generated for ease of use. The resulting model pipeline is scalable and efficient, regardless of the number of images it processes.

Additionally, to make it feel like navigating through a Google Street View of the cemetery, we had to consider GDPR and privacy concerns as there were people in the area. Therefore, once an image was received, it was split, flattened, and processed through a model that blurred out people before being sent back into the system for labeling. This anonymization was done seamlessly, and the images were stored for further labeling purposes.

Once the images were labeled, they were sent to the Optical Character Recognition (OCR) service for text detection. The images were anonymized, processed, and cropped individually before being sent to the OCR, where the results were cleaned for better information. If an image was visible from multiple sides, the information was brought together, and scripts were set up in the database to use OCR detections on all the images. Although not as complex as self-driving technology, the system required many components, and monitoring was necessary to ensure its smooth operation. Overall, the pipeline was complex but designed to run automatically and efficiently.

Hurdles and solutions

At MindTitan, we thrive on challenges, exploring the edge of innovation and crafting solutions that push boundaries. In our journey with Gratitude Service, we came across a number of hurdles that were unique and multifaceted, requiring an innovative approach. Through a series of strategic efforts, we were able to design solutions that ensured both the functionality and efficiency of our AI-driven system. Let’s delve into the challenges we encountered and the solutions we engineered.

Architecture challenges:

Our project architecture faced two significant challenges that required innovative solutions.

• Privacy Concerns

   

The first challenge centered around privacy concerns, primarily driven by the GDPR regulations and the presence of people in the cemeteries. Our priority was to ensure that the privacy rights of individuals were respected while we gathered and processed the data. To address this, we utilized specially designed AI models that could effectively anonymize images. In practice, this meant processing each image to remove or obscure identifiable features of any individuals present in the cemetery. This approach not only maintained privacy standards but also ensured our compliance with GDPR regulations. By leveraging these models, we managed to maintain the data integrity necessary for our operations without infringing on personal privacy.

• Slow Interface

   

The second challenge we faced was a slow interface, which was caused by the high-resolution image requests made by the AI model. High-resolution images, while offering more detail, also required more computational resources and time to process. This results in a sluggish interface, hampering the efficiency of the system. To tackle this automatically, we implemented scripts capable of scaling a set of images for labeling. Moreover, we converted the high-resolution images into a lower-quality format that, while still retaining essential details, allowed for faster loading and smoother operation. Through these modifications, we managed to streamline the interface while ensuring the system continued to provide accurate and meaningful results.

Data Challenges:

• Complex Data Collection

The complexity and unique features of cemetery landscapes presented us with a multitude of challenges that conventional, off-the-shelf, or synthetic data collection methods failed to meet adequately. These landscapes, marked by their intricate patterns and varying structures, generate a plethora of visual data that include multiple tricky edge cases – each requiring careful attention and analysis.

The sheer diversity and variability in tombstone designs, inscriptions, and placements, compounded by different lighting and weather conditions, make the data collection process intricate. It’s not just about getting any data; it’s about gathering high-quality, meaningful data that our AI can learn from to make accurate predictions.

Given these challenges, we have opted for a proprietary solution to collect the data needed for our operations. This solution, designed to handle the complexities inherent in our unique industry, involves a robust and comprehensive data pipeline that ensures the systematic collection, processing, and management of data.

Although implementing this solution requires significant effort and resources, the resultant data serves as a rich resource for training our AI models. These models then provide accurate and valuable insights about the cemetery landscapes and tombstone locations, driving our ambition to create the world’s largest cemetery database.

• Labeling Tombstones in Each Camera Frame

The challenge of labeling tombstones in every frame of our camera recordings is quite a monumental task. The complexity of the cemetery landscapes’ diversity, described above, multiplies when you factor in the vast quantity of frames captured during each recording session. It is not just about identifying tombstones, but also about understanding their relative positions, inscriptions, and sometimes their historical contexts.

To address this, we’ve engineered a scalable model pipeline. This pipeline is adept at handling and processing a wide range of image quantities efficiently, ensuring that no tombstone goes unlabeled, regardless of the number of frames we capture.

This solution allows us to break down the colossal task of labeling into manageable chunks. As a result, our AI models can systematically learn and predict from a comprehensive and growing database, improving their ability to recognize and understand tombstones. The scalable nature of our pipeline ensures that, as we grow and capture more images, our system can handle the increased load while maintaining the same level of efficiency and accuracy.

• Unclear engravings

One significant hurdle we encountered during our data collection and analysis process is dealing with unclear pictures of tombstones. It is a challenge to recognize engravings on images that are blurred or of poor quality due to camera movements or compression of the lighting. Pictures of tombstones which are centuries old, for example, from the eighteenth century, with their long history, tend to have engravings that are not clearly legible due to erosion and weathering over the years.

In order to effectively handle this challenge, we implemented Optical Character Recognition (OCR) solutions to improve the accuracy of text detection on tombstones. OCR is a technology that converts different types of documents, such as scanned paper documents, PDF files or images, captured by a digital camera, into editable and searchable text data. By using OCR technology, even faint and weathered engravings could often be translated into usable data.

However, OCR technology alone wasn’t enough to ensure the best quality data extraction, particularly with the level of complexity and variability presented by the old tombstones. To further refine the process, MindTitan also implemented text-cleaning scripts. These scripts worked to reduce any noise in the OCR detections, thereby enhancing the quality and accuracy of the extracted information. The result was a far cleaner, more consistent and accurate set of data derived from the engravings.

Despite our best efforts and advanced technology, it is important to understand and acknowledge the inherent limitations of the dataset. Some tombstones, due to extreme weathering or damage, may have inscriptions that are simply too faint or eroded to be accurately detected or deciphered even by humans standing in front of them. In such cases, our approach is to focus on maximizing the extraction of information from those tombstones that have readable inscriptions.

AI training process

The AI training process that we followed is essentially a human-in-the-loop system. This is a model of machine learning that involves human intervention at key stages, which helps ensure the data’s quality and accuracy, and further aids in refining the AI model.

The initial phase involves human labelers diligently annotating a given dataset. These individuals have the task of identifying and marking specific features and elements within our cemetery and tombstone data. They manually label the engravings on tombstones, noting whether there are crucial data, like names and dates. This meticulous process creates a well-annotated dataset that serves as the training ground for our AI models.

With the prepared dataset in hand, we move on to the training phase. The labeled dataset is used to teach the AI model how to recognize and interpret the information on the tombstones. This process involves feeding the data into the model, allowing it to make predictions, adjusting its parameters based on prediction errors, and repeating this until the model’s performance reaches an acceptable level. We bring human evaluators back into the loop to manually review and assess the AI model’s performance. They scrutinize the model’s results, checking its predictions against the actual data to confirm its accuracy. This human validation adds an additional layer of assurance, ensuring that the model’s performance is up to par and that it can accurately interpret the data from the tombstones.

Upon the conclusion of the training process, we validate the model’s performance. This involves testing the AI model on a separate, unseen dataset, distinct from the one used for training. The output of this test helps us gauge the model’s accuracy and reliability, as well as its ability to generalize from the training data to new data.

Conclusion

Embarking on this groundbreaking journey with Gratitude Services, MindTitan has used the power and potential of AI technology. Our collaboration serves as a benchmark, showcasing AI’s ability to bring radical transformation to even the most traditional sectors. This remarkable endeavor epitomizes an innovative approach to the cemetery services sector and illustrates AI’s broad application spectrum.

“When it comes to AI and ML in Estonia, MindTitan is typically the first partner that springs to mind. The main advantage of collaborating with MindTitan is they made us feel like we had an in-house AI team. They were an extension of our team and not just outside consultants. I am confident that I can approach them and receive a response to my query on the same day. By working closely with our business personnel, their exceptional machine learning team pinpointed the critical domains where AI could provide the most significant value. They devised a plan of action and implemented ML effectively, resulting in the maximum benefit for the business.” (Allan Selirand, Gratitude Services co-founder)

Utilizing state-of-the-art AI models, we automated the painstaking task of cemetery mapping – a procedure that previously consumed countless hours of manual labor. Our сollaboration with Gratitude Services has culminated in an innovative platform where users can conveniently locate and honor their departed loved ones, without the need for intricate location instructions.

The path to success was fraught with challenges, from handling a diverse range of data to respecting privacy concerns, and even decoding age-old engravings. However, at MindTitan, we see challenges as opportunities for growth and innovation. We designed an exclusive data collection system, built scalable models, and integrated advanced OCR solutions to overcome these hurdles. Our strategies ensured that the system functions seamlessly while adhering to the highest data quality standards and respecting privacy norms.

An integral part of our process was the human-in-the-loop system during the AI training phase. By involving human evaluators in dataset labeling, training the AI, and validating its performance, we achieved an authenticity and accuracy level crucial for a domain as sensitive as this.

With this successful partnership, Gratitude Service now stands poised to extend its services beyond Estonia. Our strategic application of AI is paving the way for their ambition of creating the world’s largest cemetery database. While there’s still work to be done and continual improvements to be made, the initial results are promising and provide a solid foundation for our mission.

Benefits of the customized AI solution by MindTitan:

• Gratitude Services received only the features they needed, tailored for their business goals
• The solution was created for the existing equipment
• Gratitude Services received the solution as complex, as required

To provide the best experience for customers, it appeared vital to choose the right AI team to collaborate with. Off-the-shelf offers were not optimized to suit the needs perfectly. Additionally, ready solutions could not provide complex and combined AI models. A custom-built solution, among many other advantages, will yield the most accurate results for your problem.

The right software partner will understand your business and technical requirements and help you figure out the best way to solve the problem.

With an AI solution set correctly in place, your business will benefit from high-accuracy results, thus improving productivity and efficiency, providing better resource and time allocation and management, and worthwhile ROIs in the long run.
This case study stands as a shining testament to the profound impact of well-applied, ethically grounded AI. It underlines the potential of AI to revolutionize industries, provide value to customers, and, ultimately, transform our interactions with the world.

MindTitan continues this exciting journey, pushing the boundaries of what’s achievable with AI. We are not just creating technology; we are creating the future.