From Implicit Ambiguity to Explicit Solidity: Diagnosing Interior Geometric Degradation in Neural Radiance Fields for Dense 3D Scene Understanding

Neural Radiance Fields (NeRFs) have emerged as a powerful paradigm for multi-view reconstruction, complementing classical photogrammetric pipelines based on Structure-from-Motion (SfM) and Multi-View Stereo (MVS). However, their reliability for quantitative 3D analysis in dense, self-occluding scenes remains poorly understood. In this study, we identify a fundamental failure mode of implicit density fields under heavy occlusion, which we term Interior Geometric Degradation (IGD). We show that transmittance-based volumetric optimization satisfies photometric supervision by reconstructing hollow or fragmented structures rather than solid interiors, leading to systematic instance undercounting. Through controlled experiments on synthetic datasets with increasing occlusion, we demonstrate that state-of-the-art mask-supervised NeRFs saturate at approximately 89% instance recovery in dense scenes, despite improved surface coherence and mask quality. To overcome this limitation, we introduce an explicit geometric pipeline based on Sparse Voxel Rasterization (SVRaster), initialized from SfM feature geometry. By projecting 2D instance masks onto an explicit voxel grid and enforcing geometric separation via recursive splitting, our approach preserves physical solidity and achieves a 95.8% recovery rate in dense clusters. A sensitivity analysis using degraded segmentation masks further shows that explicit SfM-based geometry is substantially more robust to supervision failure, recovering 43% more instances than implicit baselines. These results demonstrate that explicit geometric priors are a prerequisite for reliable quantitative analysis in highly self-occluding 3D scenes.

💡 Research Summary

This paper investigates a fundamental limitation of mask‑supervised implicit Neural Radiance Fields (NeRFs) when applied to densely packed, heavily self‑occluding scenes. The authors identify a failure mode they call Interior Geometric Degradation (IGD): because volumetric rendering relies on transmittance accumulation, rays that pass through many opaque objects lose gradient signal for interior geometry, causing the learned density field to collapse into hollow shells or fragmented structures. To quantify IGD, the study uses three synthetic fruit‑counting datasets with increasing occlusion levels (well‑separated peaches, moderately occluded apples, and densely clustered plums). Ground‑truth geometry, camera poses, and perfect binary instance masks are provided, eliminating segmentation noise.

Two state‑of‑the‑art mask‑supervised NeRF variants are evaluated: FruitNeRF, which extracts a single surface point per ray based on maximum transmittance, and InvNeRF‑Seg, which retains all sampled points along each ray and filters them by density. Both methods produce visually high‑quality renderings and accurate masks, yet on the dense plum dataset they recover only about 89 % of the true fruit instances (≈661‑662 out of 745). The authors show that improving surface coherence or mask quality does not raise this ceiling, confirming that the bottleneck is intrinsic to the implicit representation under heavy occlusion.

To overcome IGD, the paper introduces an explicit geometric pipeline based on Sparse Voxel Rasterization (SVRaster). Sparse Structure‑from‑Motion (SfM) point clouds and camera poses are first obtained with COLMAP and used to initialize a voxel grid. The 2‑D instance masks are then “lifted” onto the voxel grid by majority voting, assigning semantic labels directly to occupied voxels. This step preserves interior occupancy even when many views see only the outer surface. The voxel cloud is cleaned with color‑based background removal and density‑based outlier filtering. For instance extraction, DBSCAN clustering provides an initial grouping, and an adaptive recursive splitting procedure (K‑means with k = 2) refines clusters that exceed volume or size thresholds. Because SVRaster maintains zero‑density gaps between adjacent fruits, geometric splitting reliably separates touching objects without learned priors.

On the same plum dataset, the explicit SVRaster pipeline recovers 714 fruits, achieving a 95.8 % instance recovery rate—more than 50 additional fruits compared with the implicit baselines. The improvement is most pronounced under dense occlusion, confirming that IGD stems from representational constraints rather than dataset artifacts.

The robustness of the explicit approach is further tested with imperfect supervision. Instance masks generated by the Segment Anything Model (SAM) introduce missed detections and fragmented predictions, simulating real‑world segmentation failures. Under this degraded supervision, SVRaster still outperforms the implicit methods, recovering 43 % more instances. The authors adapt clustering parameters (larger DBSCAN epsilon, relaxed splitting thresholds) to accommodate sparser point clouds, yet the explicit geometry remains the decisive factor.

In summary, the study demonstrates that implicit NeRFs, despite their success in novel view synthesis, cannot reliably provide interior geometry for quantitative 3D analysis in highly self‑occluding scenes. Incorporating explicit geometric priors derived from SfM and leveraging a voxel‑based representation restores physical solidity, dramatically improves instance counting accuracy, and offers resilience to imperfect 2‑D supervision. The work suggests that future dense‑scene reconstruction pipelines should combine photogrammetric geometry with neural rendering to achieve both visual fidelity and quantitative reliability.