Building Camera-Dependent Picture Datasets
Greenwaves Technologies is a 5-year-old fabless semiconductor startup established in Grenoble, France. Our first product GAP8 is the world’s first IoT Application Processor armed with 8+1 RISC-V based cores and a high performance HW convolution engine. It is a simple yet very sophisticated unique processor architecture, which delivers an energy efficiency that is 20x better than the state-of-the-art, opening a large range of battery powered applications. Examples of applications are people counting, keyword spotting, combined with beamforming, object recognition, face detection and vibration analysis. GAP8 is especially effective on machine learning inference algorithms (CNN, SVM, Bayesian, Boosting, Cepstral analysis). Yet, GAP8 is by and large programmed just like a regular MCU.
Our technology is very much ahead of the state-of-art, and our chip is just about to prove its revolutionary potential on a wide open global market. For a team, it is a very motivating challenge that each of us could be part of in proportion to one’s own enthusiasm at work. As a growing and highly multicultural team with sharp personalities, Greenwaves Technologies is very proud of its specific collaborative management style. The company is and will be what we each of us make of it, as we experience every day, and we are looking for talented, enthusiastic, curious and committed people, who will be ready to bring their energy and skills for a significant contribution to the success of the company’s project.
The learning phase is key to the performances of modern artificial intelligence (AI) applications. It relies on large datasets, the quality of which (e.g. the number, quality and diversity of samples) directly impacts the quality of the results on real use-cases. As for image AI applications, (such as people counting, object recognition, face identification…), a number of open or commercial datasets are already existing, and may spare the huge effort of constructing a new one. However, image sensors used in an ultra low-power context are of lower quality. To ensure the best performances require to do the learning phase on pictures with characteristics similar to the ones of the images from those sensors.
During this internship, the candidate will study, propose and automate image processing algorithms to transform a high-quality, high-resolution dataset into a corresponding dataset as if the pictures were acquired using the camera used for the final application.
This works may require study of the camera properties (paper study and/or real-life measure), definition of the picture transforms to be applied, coding and automation of a tool able to process a dataset, experiments on a real application using the new dataset to assess performance gains.
- Digital image processing techniques (picture analyses, transforms, …);
- Programming of signal processing algorithms, preferably using C/C++ or Python;
- Software debug techniques.
- Good level of spoken and written English;
- Strong team spirit and communication abilities, in a multisite environment;
- Ability to work autonomously and proactively on assigned tasks.
- Knowledge of image sensors operation and imperfections;
- Knowledge of parallel embedded architectures and DSP processors;
- Familiarity with versioning/revision control systems (e.g. git).
- Engineer or Master’s degree student in digital signal processing, preferably with an image processing specialization.
Employment type: Internship
Location: Grenoble, France (Alsace-Lorraine near train/tram)