Empirically verify Reliev3R’s behavior under large-scale data scaling

Determine the empirical behavior of Reliev3R when its training is scaled to substantially larger datasets, in order to assess whether the weakly supervised paradigm maintains or improves reconstruction and camera pose estimation performance without multi-view geometric annotations.

Background

Reliev3R is proposed as a weakly supervised training paradigm for feed-forward 3D reconstruction models that avoids reliance on multi-view geometric annotations by using pseudo monocular relative depth and sparse image correspondences. A central motivation is that removing costly SfM/MVS labels could enable scaling to larger, more diverse training data.

However, the paper’s experiments are conducted on DL3DV-10K-scale data, and the authors explicitly state that they have not empirically validated Reliev3R’s behavior at substantially larger scales. This leaves open whether the intended data-scaling advantages translate into measurable performance gains or robustness improvements when training is scaled up.