Synthetic Data Generation
We can generate photo-realistic data in Virtual Reality
Good deep learning models need thousands of samples to train on. In many applications, it can be difficult to collect a balanced training set, because some classes are represented by only a few samples. For instance, in applications used for industrial quality control some failure modes occur in the range of one in a million, making it almost impossible to collect a thousand samples from that failure mode. In other cases, the cost of collecting data is prohibitive. One example is reinforcement learning models in robotics: collecting samples for bad outcomes (i.e. cases in which the robot might break itself or something else) can be very costly.
One effective way to compensate for the lack of real data is to synthesize data in virtual reality simulations. We use Unreal Engine to generate photorealistic samples in a wide variety of scenes as viewed from various angles under different lighting conditions. We can generate 2D or 3D images or videos with visible surface depth maps and with occlusions that are correctly accounted for. Since the per-pixel ground truth data is readily available in the simulation, there is no need to spend time and money on the annotation.
Annotation
We can speed up your data annotation process.
No AI solution can work without high-quality data annotation, i.e., without adding labels, masks, and numeric metadata to the training data. Annotation is often the bottleneck in an automation project because it may require 80% to 90% of the total human effort devoted to the project.
We do it differently.
After a small portion of the data has been manually annotated, the MInD Annotator learns to semi-automatically annotate the data. Next, the platform recommends annotations that humans can revise, and in this way, the pre-annotating function can be iteratively improved. The result is an annotation process that is 10 to 100 times faster than a manual process.
Consultancy
Both at the Engineering and the Business level
Engineering
At the core of our business is our experience with the practical applications of deep learning. In our consultation projects, we work closely with you to solve engineering and business problems.
- We plan the automation process with you.
- We help design the data collection.
- We can either annotate the data for you or train your staff to annotate the data.
- We train your engineers to deliver AI solutions and to maintain them on their own.
- We take the responsibility of specifying all of your hardware needs, and we connect you with our partners who will install and maintain the hardware.
Business
We can also assess your business from a broader perspective. We help business managers and executives consider how you can employ AI across your company to expand your business. We help you to:
- Identify jobs that can be automated.
- Reengineer business processes to cut costs or speed up your services.
MI-Crowd
Real time crowd counting on security cameras
MI-Crowd is an out-of-the box software product for real-time crowd counting. It is recommended for counting crowds up to 200 people, i.e. it is an ideal solution for public transport, conference halls or other public spaces.
Advantages
It has numerous advantages over classic people counting systems placed above the entrances:
- It can count the crowd based on a single image, thus the estimation error can be diversified away by evaluating multiple images. E.g. taking multiple images between two bus stops can increase the 94% single-image accuracy up to 98%.
- It counts only the heads, so other objects (e.g. luggage, umbrella) that can mislead the classic counters are ignored.
- It is ready to be used on any camera using arbitrary image resolution.
Technical information
The inference speed of the model using 640×480 image resolution were tested on two hardver configurations:
- GPU (GTX3070): 60 ms, i.e. about 17 images per second
- CPU (i9): 400 ms, i.e. 2.5 images per second