NVRQ: A Benchmark for Object-Level Quality in 3D Neural Reconstruction

Wangren Xu, Bidur Khanal, Naveen Kumar Rai, Bjoern Haefner, Sean M. Pieper, Nick Schneider

Keywords: benchmark, novel view reconstruction quality assessment

Abstract:

Realistic real-world sensor simulations are key to the development of modern end-to-end models for Physical AI. Neural reconstruction has emerged as a promising solution, but it often shows degradations when rendering novel views.
Evaluating the quality of novel views remains challenging because standard scene-level image quality metrics often average away localized artifacts on safety-critical actors such as pedestrians, cyclists, and vehicles.
We address this gap by introducing a framework for object level evaluation that establishes strict 3D-to-2D correspondences between original and novel rendered views which are then used to compute semantic similarity metric.
Additionally, we're publishing a dataset for novel view synthesis quality evaluation based on the Physical AI NuRec dataset. The dataset contains pre-reconstructed 3D driving scenes, novel rendered views, corresponding 3D semantic bounding boxes and subjective human quality ratings.
Experiments show that our Novel View Reconstruction Quality (NVRQ) metric consistently outperforms both classical image quality metrics and recent Novel View Synthesis (NVS)-specific baselines, achieving the strongest correlation with downstream task degradation and human judgments.

Submission Number: 19