Neural Density-Distance Fields

1 Introduction

• NeDDF reciprocally constrains distance and density fields to support robust localization while retaining reconstruction quality comparable to NeRF.The representation is designed to combine distance-field registration with density-field expressiveness.Source: §1 · pp. 1–2
• Density fields are expressive, but their gradients are unavailable across most regions, whereas distance fields provide gradients and residuals useful for registration.The introduction notes that this gradient availability can support more convex registration objectives.Source: §1 · p. 1, Figure 2 · p. 3
• Conventional unsigned distance fields fail on ambiguous hairball density changes and low-density glass, cases that NeDDF represents properly.The figure contrasts iron, hair, and glass examples across the fields.Source: Figure 1 · p. 2
• NeDDF extends distance fields to arbitrary density distributions through explicit distance-to-density conversion, avoiding density constraints tied to explicit boundaries.It interprets distance using depth derived from volume rendering and fits translucent-object density to mid-level gradient magnitudes.Source: §1 · p. 2, Figure 3 · p. 3
• The paper's contributions include recovering density from independent points using distance and gradient information, and alleviating distance-gradient instability from cusps and sampling frequency.The authors evaluate expressiveness and registration performance experimentally.Source: §1 · p. 2

2.1 Neural Fields

• Neural fields represent continuous signals with neural networks and can encode high-frequency shapes with fewer parameters than conventional discrete representations.Voxel complexity grows cubically with resolution, while grid structures are limited by Nyquist frequency and point-based representations interpolate discrete coordinates.Source: §2.1 · p. 3
• Neural fields can be treated as continuously differentiable, and NeDDF explicitly couples distance and density so each can use its appropriate objective without a consistency penalty.The distance gradient describes density information in the proposed model.Source: §2.1 · p. 4

2.2 Density Field

• Density fields represent translucent and high-frequency scenes, but uncertain gradients in empty regions limit localization; NeDDF supplies a consistent distance field and gradients for rough-initial-pose registration.Some existing NeRF-based methods additionally require close or constrained camera configurations, optical flow, or particular image conditions.Source: §2.2 · p. 4

2.3 Distance Field

• Distance fields provide residuals and gradients for registration, while prior neural distance-field methods include assumptions of boundary surfaces that limit representable shapes.NeDDF extends distance fields to varied density distributions using depth values derived from volume rendering.Source: §2.3 · pp. 4–5

3.1 Distance Field from Density Field

• NeDDF interprets density-derived volume-rendering depth as distance, taking its shortest value across viewing directions to remove view dependence.The minimizing direction can be discontinuous, complicating this calculation.Source: §3.1 · pp. 5–6

3.2 Density from Distance Field

• The distance derivative and gradient magnitude yield density at independent points, using sufficiently small t_n to address undefined density near D≈0.The conversion is differentiable and enables mutually constrained distance and density fields.Source: §3.2 · pp. 6–7

3.3 Removing Cusps

• A 4D auxiliary gradient redistributes gradients near nondifferentiable cusps to suppress spurious densities, while a penalty constrains its global shape.The heuristic constraint is applied when ∂D/∂t > 0.Source: §3.3 · p. 7

• The blank-region penalty discourages color changes along the distance-gradient direction, making the nearest neighboring object's color, direction, and distance available from outside samples.The penalty is active outside the object because the distance-gradient magnitude is small inside it.Source: §4 · pp. 7–8
• NeDDF estimates camera pose by projecting weighted pseudo-correspondence points synthesized from near-surface samples and minimizing their distance from the corresponding image pixels.For each ray, samples closer in color to the observed than rendered color are combined, with weights favoring similar colors or smaller distances.Source: §4 · p. 8, Figure 8 · p. 8
• A sample at position p_i yields a near-surface point at p_i − D_i∇D_i, using its predicted distance and distance gradient.The network also supplies the sample color, distance, and gradient; distance-gradient direction is used to regularize color in blank regions.Source: §4 · p. 8, Figure 8 · p. 8
• The reprojection error is the distance between a ray's pixel coordinates and the projection of its pseudo-correspondence point.Source: §4 · p. 8

5.1.1 (a) Quality of novel view synthesis.

• The experiments compare NeDDF with NeRF and NeuS on PSNR using eight synthetic scenes, and test pose localization across 200 Lego-scene viewpoints with randomized initial rotations and translations.Pose comparisons include photometric-only, reprojection-only, and reprojection followed by photometric optimization.Source: §5.1.1 · p. 8, §5.1.2 · p. 8

5.2.1 (a) Quality of novel view synthesis.

• NeDDF achieves novel-view quality comparable to state-of-the-art methods, although its PSNR is slightly inferior; it also reconstructs transparent drum parts and delicate leaves that NeuS handles poorly.The authors attribute the thin-surface reconstruction to representing occupancy below the sampling interval as lower density.Source: §5.2.1 · p. 9, Table 1 · p. 9, Figure 9 · p. 9

5.2.2 (b) Accuracy of localization.

• Reprojection-error pose optimization followed by photometric optimization significantly expands the range of initial poses yielding highly accurate localization.Photometric-only optimization is accurate around initial position errors of 0.5 m or less and angular errors of 10 deg or less, but errors can grow from larger initial errors; reprojection alone gives limited residual improvement.Source: §5.2.2 · pp. 9–10, Figure 10 · p. 9
• With large initial pose errors, reprojection alignment makes photometric optimization effective by iteration 100, whereas photometric-only optimization makes little progress.Some color-based correspondences are mismatched; the authors suggest reducing their weights or using more distinctive features.Source: §5.2.2 · p. 10, Figure 11 · p. 10

• NeDDF explicitly converts between distance and density using distance and gradient information, while an auxiliary gradient addresses discontinuity points.Source: §6 · p. 10
• Experiments show competitive novel-view quality, more stable meshes, and a more comprehensive range of camera poses followed than NeRF.Source: §6 · p. 10
• Because NeDDF uses an unsigned distance field, it lacks helpful distance-gradient directions inside objects and retains NeRF's time-consuming optimization and rendering.The authors suggest that recent NeRF speedup and stabilization advances may also apply to NeDDF.Source: §6 · p. 10

• Density is calculated from the first derivative of the distance output, requiring activation functions with second derivatives; the architecture uses tanhExp, whose second derivative is continuous.Source: §7 · p. 13
• The architecture damps high-frequency positional-encoding components so frequency scales in the single-derivative hierarchy match those of conventional encoding.The conventional encoding's frequency range otherwise creates a scale difference of 2^(L−1) in its influence on the density derivative, making learning unstable.Source: §7 · p. 13
• Conventional positional encoding is added after the distance-output layers to recover detailed color fields, but high-frequency color restoration is worse than NeRF at the same network size.Source: §7 · p. 13

8 Deriving the conversion from distance to density

• The derivation considers D(r(t)) along r(t) = p + tv, slicing the distance field in gradient direction v.It then calculates the derivative in that direction.Source: §8 · p. 13
• The derivation treats the displayed equations as valid when t_f is sufficiently large for all light to be reflected.Source: §8 · p. 14
• The derivation assumes t_n is small enough that T(t_n) = 1 and takes the limit as ∆t approaches 0.Under this assumption, the cited expression converges to ∆tσ(r(t_n))t_n.Source: §8 · p. 15
• The final displayed relation includes the factor (D(r(0)) − t_n)σ(r(t_n)) minus 1.Source: §8 · p. 15

9 Parameter selection for the shape of the auxiliary gradient

• For α ≤1, the auxiliary gradient is active over a narrower range than the distance field, while larger α concentrates it nearer cusps.The constrained shape yields a unique set of auxiliary gradients.Source: §9 · p. 15, Figure 13 · p. 15
• The auxiliary gradient activates strongly near distance-field cusps, where distances from several objects are similar.Source: §9 · p. 15
• Without auxiliary gradients, incorrect volume densities occur in empty regions.The passage compares rendering results with and without the gradients.Source: §9 · p. 15

10 Evaluation of reconstruction performance in smoke scenes

• The smoke-scene dataset has 100 hemispherical training and validation viewpoints and 200 orbit test viewpoints, with RGB and transmittance recorded at 800 × 800 resolution.Source: §10 · p. 15
• NeDDF produces high-quality novel views in smoke scenes, with a distance field that increases inside the object and a density field representing translucency.The distance minima are larger than in Fig. 14, consistent with the stated assumption that low density corresponds to larger distance minima.Source: §10 · p. 15, Figure 16 · p. 17, Figure 17 · p. 18

11 Evaluation of localization performance in other scenes

• Reprojection error improves camera-localization performance over photometric error alone across the evaluated scenes, but can degrade or fail to improve results in some cases.Materials may suffer from correspondence mismatches, while Mic and Ship retain photometric local solutions.Source: §11 · p. 16, Figure 18 · p. 18