Figure 2.

EndoCV2021 challenge tasks: participants performed model training on white light imaging data collected from five centres (C1–C5). The tasks included detection and segmentation. Trained models were then tested on both seen and unseen centre datasets and on unseen data modality (widely used narrow-band imaging). Generalisability assessment is obtained by computing deviations between these unseen samples w.r.t. seen samples. Task outputs included bounding box prediction with confidence and class label for detection task and binary mask prediction for polyp segmentation.