Appropriate depth range for cost-volume construction on PASCAL-Context

Determine the appropriate depth range for sampling the candidate depth planes used to construct the cost volume in the Cross-view Module when applying the 3D-aware multi-task learning framework to the PASCAL-Context dataset, which lacks ground-truth depth annotations, in order to ensure effective cross-view geometric correlation modeling.

Background

The proposed framework augments a multi-task learning encoder with a Cross-view Module that builds a depth-parameterized cost volume to encode cross-view geometric correlations. Constructing this cost volume requires choosing a depth range over which candidate planes are sampled.

While NYUv2 provides depth labels and a known 0–10 m range, PASCAL-Context has no ground-truth depth, making the choice of depth range nontrivial. The paper notes this uncertainty and explores several ranges empirically, but a principled determination of the appropriate depth range remains unresolved.