A Unified Evaluation Protocol and Late-Fusion System for Monocular Per-Object Distance Estimation in Indoor Scenes

Monocular per-object distance estimation aims to predict a metric distance for each detected object from a single RGB image. Although monocular dense depth estimation and monocular 3D object detection are well studied, indoor object-level distance estimation remains weakly standardized. In dense depth-based pipelines, object distance is commonly obtained by combining object detection with depth prediction and aggregating depth values within a region of interest. However, existing approaches differ in region selection, detection source, and aggregation strategy, limiting comparability across methods. This paper proposes a unified, deployment-aligned evaluation protocol for dense depth-based pipelines, where both predicted and reference distances are computed within the same predicted bounding box. This formulation standardizes region-of-interest selection, removes dependence on ground-truth boxes at inference time, and enables consistent evaluation across late-fusion methods. The framework integrates a curriculum-trained Depth Anything V2 ViT-S model for metric depth estimation with a YOLO11n detector for object localization. Under the same-box evaluation protocol on SUN RGB-D, the Depth Anything V2 ViT-S backbone achieves object-wise distance estimation accuracy of MAE = 0.1286 m, RMSE = 0.1817 m, AbsRel = 0.0662, and δ1 = 0.9785 using mean aggregation over valid box depths. Scaling the backbone from ViT-S to ViT-L further improves performance to MAE = 0.1059 m, RMSE = 0.1484 m, AbsRel = 0.0553, and δ1 = 0.9898, corresponding to approximately 17.7% lower MAE and 18.3% lower RMSE relative to ViT-S, and provides a standardized reference point for indoor object-level distance evaluation under the same-box protocol.