Origin of durian’s robust segmentation under viewpoint changes

Investigate whether the stable segmentation performance of the MVImgNet durian class under viewpoint shifts is primarily driven by distinctive visual features of the durian itself or by contextual cues in the surrounding scene (e.g., background elements such as curtains).

Background

The durian class shows unusually strong and consistent segmentation performance across angles, despite its nearly spherical geometry. The authors explicitly note uncertainty about whether this robustness is due to the object’s visual signature or scene context.

Clarifying the source of robustness would help distinguish object-centric from context-centric generalization and guide representation learning for multi-view segmentation.

References

It remains unclear whether the stable segmentation under viewpoint shifts stems from its distinctive visual signature or from contextual cues in the surrounding scene, such as the curtains highlighted in autoref{fig:all_angles_objects}.

— Evaluating Foundation Models' 3D Understanding Through Multi-View Correspondence Analysis (2512.11574 - Lilova et al., 12 Dec 2025) in Appendix A, Experiment A: durian figure caption