Synthetic Data and AI for Good: Utilising Computer Vision for Accurate Mine Detection

syntheticAIdata collaborated with the DEMINE Foundation to address their challenge of acquiring high-quality data for computer vision model training. We donated a dataset of realistic synthetic images representing various scenarios of landmines and explosive devices within authentic environments to the DEMINE Foundation's project.

Client

DEMINE Foundation

Industry

AI for Good

Location

United Kingdom, EU

Platform

NVIDIA Jetson, DJI drones

DEMINE Foundation saving lives through AI innovation

Customer

DEMINE Foundation is a volunteer non-profit organization aiming to create and deliver low-cost mine detection and landmine clearance solutions using artificial intelligence and off-the-shelf drones, sensors, and computers.

Swarm-capable drones are equipped with multiple cameras and sensors. This multi-modal approach enhances the platforms' perception capabilities, allowing for a more comprehensive data collection and analysis.

Real-time object detection and recognition are enabled with the use of NVIDIA Jetson Orin, a powerful computing system designed for advanced robotics and AI applications. Video feeds from the drones are streamed to the NVIDIA Jetson device and analysed in the real-time.

Website: www.deminefoundation.com

Challenge

Gathering data on landmines and unexploded ordnance for computer vision training presents several challenges. These challenges are primarily due to the hazardous nature of the devices and the inaccessible locations where they are often found. The presence of landmines in conflict zones, post-war areas, or remote regions makes it dangerous for humans to approach and collect data manually. This limits the availability of ground truth data necessary for training and validating computer vision models. The variability in terrain and conditions adds complexity to the task. It requires innovative solutions and technologies to address this critical issue effectively.

3D render of a terrain covered with mines

chevron_left

chevron_right

Solution

syntheticAIdata Enterprise provides many advanced features for synthetic data generation that are suitable for this use case.

Our platform generated synthetic images that simulate a variety of landmine and explosive device scenarios within authentic environments. These images were created to represent mines in different landscapes, complete with a diverse selection of natural vegetation and foliage.

The level of detail on the synthetic images ensures that models can develop a deep understanding of the complexities involved in identifying landmines within diverse environmental contexts.

Result

Utilising synthetic data significantly enhanced the AI model's accuracy by 20% and improved its performance across all metrics. Additionally, the incorporation of synthetic data reduced false positives by over two times.

20%

Improved accuracy

Less false positives

We strive to use Machine Learning and AI to save lives. Thanks to the data generated by syntheticAIdata we were able to increase the accuracy of our model by 20%. Moreover, incorporating generated synthetic data into the training of our model helped reduce the number of false positives by more than two times.

Pavlo Melnyk, Co-Founder and CTO of DEMINE Foundation