The goal of Scortex is to be able to deploy quickly its solution for any new client or reference. Since Scortex technology currently relies mostly on supervised algorithms, an extensive manual annotation process is required. State of the art shows that it is possible to alleviate the need for annotations through several techniques. One of these is the use of unsupervised / semi supervised and weakly supervised learning.
The goal of the thesis is thus to design and implement a fast, non purely supervised defect detection system on high resolution images and videos. The thesis will typically include the study of deep anomaly detection systems. The candidate may have to explore several leads from the bibliography such as auto encoders, GANS, metric learning and siamese networks or more traditional methods for anomaly detection (LOF, One class SVM)...
Scortex is looking for a way to proof its supervised algorithms and to be able to generate smart candidates for labelling. As a result, the anomaly detection system should ideally give a segmentation mask of the defect, frame by frame.
As a proactive member of the machine learning and computer vision team, your work will include a varied range of challenges:
- explore various state of the art techniques to help solve tasks currently unbeaten by computers;
- stay on the bleeding edge of research and participate actively in the community;
- design, develop and implement supervised and unsupervised models with extremely constraining requirements not only on accuracy, but also on real-time execution, fast and scalable training processes and minimal annotation levels;
- help improve our pipelines of data acquisition, training and inference.