Example application
Autonomous systems
Robots and drones can use Artificial Intelligence for environment segmentation, obstacle avoidance and object interaction.
Task: Store sensor data that has been recorded by the system. This could be camera images, radar samples, laser or structured light. The amount of data needed depends on the complexity of the robots environment and detail level needed for the application. Count on amounts ranging from 1000 to 10000 samples.
Tools: We have expertise to utilize cloud service providers such as AWS to implement solutions that that matches your storage needs. Have a look at https://aws.amazon.com/s3/ for more information on cloud storage.
Cost: Storage cost depends on your needs, have a look at https://calculator.aws/#/addService, search for s3. Count on amounts in the range of 30$ a month.
Time: It takes days to setup cloud services for storing data.
Task: Annotate the data so that every sample represents something meaningful.
Tools: We have expertise with tools that simplifies this process. It can also be outsourced with services such as https://www.mturk.com/. Also open source datasets that are already annotated can be utilized to solve parts of your problems. Have a look at https://storage.googleapis.com/openimages/web/index.html to browse for data that could suite your domain.
Cost: All the tools are free of charge.
Time: It takes weeks to annotate thousands of images, this can be parallelized by multiple workers.
Task: Find a machine learning model that can be used by your system. Real-time requirements, accuracy requirements and available compute capacity are guidelines when choosing the appropriate model.
Tools: We have expertise to use frameworks and models that can be picked off the shelf that matches your needs. Have a look at https://keras.io/api/applications/, https://modelzoo.co/ and https://gluon-cv.mxnet.io/model_zoo/index.html.
Cost: All the tools are free of charge.
Time: It takes somewhere between a week and weeks to choose an appropriate model.
Task: Train a machine learning model using your data.
Tools: We have expertise to train machine learning models with open source frameworks and services. Have a look at https://www.tensorflow.org/, https://pytorch.org/ and https://keras.io/. Also have a look at services like https://databricks.com/ for clustered computing that integrates with your cloud storage.
Cost: Minimum cost for a local server that can be used to train models can be acquired for around 2500$. Using managed cloud service such as DataBricks will cost you about 120$ per month.
Time: It takes weeks of experimentation to train a model.
Task: Optimize model for use in real-time application.
Tools: We have expertise to optimize your models for demanding target systems. The appropriate optimization tool depends on the choice of training framework and processing hardware on your target platform. Have a look at https://www.tensorflow.org/lite or https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/graph_transforms/README.md or https://docs.openvinotoolkit.org/latest/index.html.
Cost: All the tools are free of charge.
Time: It takes a week to create an optimized model.
Task: Create application that uses model to make predictions based on sensor data.
Tools: We have expertise to create applications in C/C++ that integrates the optimized models to solve problems in real time using public frameworks. Look at https://github.com/opencv/opencv/tree/master/samples/dnn, https://developer.arm.com/ip-products/processors/machine-learning/arm-nn and https://docs.openvinotoolkit.org/latest/index.html.
Cost: All the tools we need are free of charge.
Time: It takes weeks to create an application that is integrated into your system.
