Figure 4

Qualitative results. Top row: False positive model predictions (OoBNet predicts the frame to be out-of-body even though it is not). Bottom row: False negative model predictions (OoBNet predicts the frame to be inside the body even though it is out-of-body). Below each image the binary human ground truth annotations and the probability like model predictions are provided. In (A), surgical smoke is impairing the vision. In (B–D), a mesh, a swab, and tissue are so close, that—lacking the temporal context—it is difficult to distinguish even for a human annotator whether it is out-of-body or not. In (E) and (F), blood on the endoscope and a glove with bloodstains mimic an inside view. In (G), a surgical towel covers most of the patient’s body, so that the model lacks visual cues for an out-of-body frame. In (H), the endoscope is cleaned in a thermos, which mimics the inside of a metal trocar.