Ensemble-based Endoscopy Artefact Detection

Proposed an ensemble of RetinaNet-based object detectors to localize bounding boxes and predict labels of eight different artefact classes that generalizes to an inter-patient, multi-tissue and a multi-modal corpus of endoscopy video frame data

The common artefacts of interest that corrupt endoscopy video frames include contrast, saturation, instrument, blood, specularity, blur, imaging artefacts and bubbles.

Achieved an mAP of 0.3405 improving existing state-of-the-art results

Advisor: Leela Velusamy