Visual Pipeline Optimisation

When operating on unlabelled data, techniques have to be engineered to rapidly bootstrap the training of networks and the detection of outliers as the labelling is gradually performed. The processing of input images requires more than naive training with manual ground truth: anomaly detection, geometric rectification, and dimensionality reduction techniques all have a roll to play.

We are working with data acquired during public-health surveys of Covid-19 antibodies, captured in the form of photographs. This project aims to reduce the computational cost of certain portions of the pipeline, and to perform extra analyses, beyond outcome detection, which includes sub-threshold immune response, and rejecting poor quality images or segmentation.

Though this may sound like a typical machine learning/computer vision project, it isn’t! Live returns of image samples are typically over 100,000 in size, are truly “in the wild”, and uncurated. The system has to turn around the data and produce analytical information within a short time-frame. If you are interested in this sort of challenge, which includes custom data science, engineering and computer vision, then drop us a line.