Kognic - Master thesis proposal

Intelligent Quality Assurance Process

About Kognic
Kognic provides the only software available in the market tailored to measure and improve perception performance for autonomous mobility. Our solutions empower engineers and product teams to argue safety in an objective and data driven way while helping balancing needs across departments.

We provide our customers in the automotive industry with the platform to develop and validate their safe perception system. With it, they can develop, validate and use safe perception systems right from their sensor data in a scalable and cost-efficient manner. Right now, our software divides into our Ground Truth Platform and our Dataset Quality Analytics, but we will soon unlock additional products as part of our Perception Analytics.

Our goal is to create great automated driving experiences without accidents through safe perception. Our mission: to make cars pass their vision test.

Problem
Kognic provides a unique platform for ground truth annotation. One of our main commitments to our customers is about the data quality we deliver. As of today, our data quality process relies on a mix of manual and automated work.

Kognic is interested in developing innovative quality assurance tools based on artificial intelligence and machine learning. Some initial work has been carried out on classification tasks and object detection in order to detect the annotation mistakes in a dataset, but a lot of efforts and challenges remain:

One is about the amount of paradigms we must handle. Our clients will contact us with requests to annotate and quality assure data coming from cameras, radars, lidars, IMU, etc. These sensors also have different technical specifications: resolution, point density, effective range, field of view, etc. In addition, we support a multitude of annotation types: 2d boxes, 3D cuboids, 2D/3D semantic segmentation, curves, etc. We also need to track semantic properties, such as an object's colour, shape, state, etc.

In order to be able to perform automatic data quality checks, a large amount of AI algorithms must be developed. They must support different kinds of inputs and be able to adapt to the complexity and subtleties of the detection tasks.

The first topic of interest for us is to improve our existing workflow and develop new automatic quality tools. The second is to understand how the difference in the annotation task (sensor, annotation type) can impact the quality process.

We are mainly interested in exploring solutions for quality assurance for 2 tasks:

- 3D data (lidar)

- Sequence of frames (2D or 3D)



Expected work

- Perform a thorough literature study and select one or two suitable solutions to perform quality assurance

- Participate in the selection and preprocessing of required data necessary for the solution

- Implement and validate the selected solutions

- Suggest an effective way to implement the developed solution to Kognics' platform



Reference / Links
1] Measuring data quality efficiently (https://www.kognic.com/articles/measuring-data-quality-efficiently/)

2] Dataset Quality Analytics (https://www.kognic.com/articles/understand-dataset-quality-through-dataset-quality-analytics-product/)

3] Why (even random) annotation errors are problematic for perception (https://www.kognic.com/articles/why-even-random-annotation-errors-are-problematic-for-perception/)