Disambiguate Network Predictions from Human Orientation Labels

Determine whether SonarVoxNet’s predicted full 3D orientation for a non-rigid, freely posed diver or the corresponding human-provided orientation label is more accurate in cases where the two disagree.

Background

Diver3D provides full-SO(3) orientation annotations for divers whose bodies are non-rigid and can assume freely varying poses underwater. Such annotations are intrinsically difficult to make exact because sparse and noisy 3D sonar returns may not uniquely determine the diver’s complete body extent or orientation. The paper reports cases in which the detector’s predicted orientation and the human annotation differ, but does not establish which is correct, leaving the reliability of the prediction-versus-label comparison unresolved.

References

Third, full-rotation annotation for a non-rigid, freely posed diver is inherently harder to make exact, and we encountered cases where it was unclear whether the network prediction or the human label was more accurate.

— SonarVoxNet: Diver Detection in 3D Bounding Box using 3D Sonar  (2610.01644 - Park et al., 1 Oct 2026) in Section 6.1, “Limitations and Future Work” (subsection labeled ssec:limitations)