Abstract
The Endoscopic Artefact Detection challenge comprises tasks for detection and segmentation of artefacts found in endoscopic imaging, with a specific task for evaluating the generalization capacity of detection algorithms on external data. For the detection of artefacts, we train RetinaNet and FasterRCNN models. To segment artefacts from the endoscopic images, we train a Deeplab v3 model and a U-Net model and also implement post-processing techniques such as the usage of an EAST text detector for detection of text artefacts and pixel-wise voting ensemble after applying test time augmentation. We observe that the RetinaNet model with a ResNet101 feature extractor is the best performing model across all object detection tasks, while the U-Net performs best in the segmentation tasks. We also implement a model agnostic object tracking pipeline utilizing image correlation-based trackers to reduce the inference time of object detection models. We believe that this pipeline can enable real-time analysis of endoscopic images in systems with processing constraints.
