Analysis of object detection performance according to implicit and explicit representations in novel view synthesis

Novel View Synthesis (NVS) has demonstrated significant potential for 3D object detection. NVS methods are categorized into implicit and explicit representations, which exhibit distinct characteristics in scene description and boundary delineation. However, directly comparing these methods within a unified detection framework is challenging due to the inherent heterogeneity of their data structures. To address this, we propose a comparative evaluation framework that analyzes the correlation between pixel-level rendering quality and object-level detection ability by performing detection on rendered 2D images. We conduct a comprehensive case study using implicit and explicit methods to investigate how their representational differences influence detection performance. Our experiments reveal that 3D Gaussian Splatting generally offers superior detection ability, primarily due to its explicit nature that preserves sharp object boundaries. Furthermore, we demonstrate that while perceptual metrics like LPIPS correlate with detection accuracy, the capability to maintain structural integrity is the decisive factor for machine perception.