acceptodds
Under review as a conference paper at ICLR 2027

Light Field Primitive for Novel View Synthesis

Abstract

We present ight ield rimitives (), a formulation for novel view synthesis that replaces the dense ray database with a compact set of differentiable primitives in the classical two-plane parameterization. Each primitive condenses a group of rays into one learned record, and its response to a query is governed by how closely that query belongs to the group. Rendering a camera ray then reduces to compositing all responses it elicits, and a scene can be optimized directly from posed images and rendered in real time with rays. Beyond its competitive performance on standard benchmarks, the main advantage of LFP is structural: its primitives reside directly in the 4D ray space, so optical and appearance effects that are already operations on the light field become behaviors of a single shared renderer. With minimal changes to that renderer, LFP supports multi-scale anti-aliasing, defocus deblurring with refocusing, rendering for fisheye cameras, and even transparent object reconstruction with ray refraction, matching specialized frameworks that devote substantial machinery to these effects.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Light field primitive, novel view, anti-aliasing, fisheye camera

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_6Z2hOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Light Field Primitives (LFP), a novel view synthesis representation that enables explicit differentiable primitive rendering into 4D ray space. Rather than representing scenes with 3D geometry like 3DGS, LFP represents a scene as a collection of ray-space primitives. Each primitive corresponds to a disparity-consistent ray group defined using the classical two-plane parameterization. The primitive stores appearance, opacity, and view-dependent radiance parameters for the ray group.

The key idea is that rays that observe the same scene content satisfy an affine coupling determined by disparity and can therefore be grouped together. Rendering is performed directly in ray space by measuring the deviation between a query ray and primitives and compositing primitive contributions in a way similar to point-based rendering.

Beyond standard NVS, several image formation phenomena are naturally ray-space operations. Consequently, multi-scale anti-aliasing, defocus deblurring with refocusing, fisheye rendering, and transparent-object reconstruction can all be supported using the same renderer with minimal modifications.

Experiments show competitive or state-of-the-art performance on standard NVS benchmarks while maintaining real-time rendering. The representation performs competitively on several tasks that typically require dedicated rendering approach. This validates the paper's central claim.

Critical strengths
  1. Novel and well-motivated representation paradigm

The paper proposes an explicit primitive representation in 4D ray space, filling a conceptual gap between implicit ray-space methods (e.g., Neural Light Field) and explicit geometry-based methods (e.g., 3DGS). The primitive formulation is technically coherent framework, that combines components such as epipolar geometry, opacity compositing, and view-dependent appearance modeling, etc. This is a good conceptual contribution.

  1. Strong validation across multiple tasks

The experimental section goes beyond standard NVS benchmarks and evaluates anti-aliasing, defocus deblurring and refocusing, fisheye rendering, and transparent-object reconstruction. These experiments directly validate the central claim that many imaging effects become natural operations in ray space.

The fact that the same renderer can be reused across all settings with minimal modifications is compelling. The transparent-object and fisheye experiments are scenarios that are ray-centric and often require substantial task-specific engineering for geometry-based representations. The ablation study further strengthens the paper. It provides convincing empirical support for the design decisions.

Critical weaknesses
  1. Insufficient discussion/comparison with respect to Neural Light Field

While the paper extensively compares and contrasts itself with NeRF and 3DGS, its discussion of prior ray-space approaches is surprisingly brief. In particular, the relationship to neural light-field methods such as Learning Neural Light Fields with Ray-Space Embedding is underdeveloped.

Given that both approaches operate directly in ray space, a more thorough comparison is necessary to better position the paper. While I believe LFP is substantially different from prior neural light fields, the paper does not make this distinction as explicitly as it should.

Questions for authors
  1. How does primitive count and memory consumption vary across scene categories? Are there systematic differences between predominantly diffuse scenes and strongly specular or refractive scenes?

  2. The atlas ablation convincingly justifies the use of multiple atlases. Could the authors comment on scenarios where the training-camera distribution provides poor coverage of scene content? Are failures dominated by insufficient capture coverage or by the atlas parameterization?

  3. Have the authors evaluated sparse-view settings? Since the representation organizes primitives according to ray coherence rather than explicit geometry, it would be useful to understand how the method behaves when viewpoint sampling becomes sparser.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 6Z2h

We thank the reviewer for the positive assessment and the insightful questions.

W1. Relation to neural light fields

We added a detailed paragraph to the related work (Sec. A). The differences are:

  1. Implicit vs. explicit. Neural light fields represent radiance with a network defined over ray space, whereas LFP uses explicit primitives, each associated with one ray group.
  2. Learned vs. built-in epipolar structure. Neural light fields can encode epipolar coherence implicitly, for example through the learned ray-space embedding of Attal et al. (2022). In LFP, the epipolar coupling is built into each primitive through its explicit disparity, and occlusion is handled by front-to-back compositing.
  3. Rendering and edits. Neural light fields evaluate a network for each query ray, whereas LFP is rasterized in real time. Anti-aliasing and defocus are closed-form modifications of each primitive's covariance Σ_i (Sec. D), a direct per-primitive control that is not readily available in a neural field.

Q1. Primitive count and memory by category (new Tab. 15)

Category (#scenes)LFP #prims / MB3DGS #prims / MB
Mip-NeRF 360 outdoor (5)2.77 M / 4154.84 M / 1145
Mip-NeRF 360 indoor (4)1.09 M / 1641.45 M / 344
Tanks and Temples (21)1.00 M / 1501.77 M / 418
Blender (8)0.42 M / 630.29 M / 68
Refractive (4)1.01 M / 1511.95 M / 461

For both methods, the primitive count varies across categories, with the largest counts for unbounded outdoor scenes and the smallest for synthetic objects. LFP is more compact than 3DGS on every real-scene category and on the refractive scenes, where the gap is largest. On the synthetic Blender scenes, LFP uses more primitives but requires a similar amount of memory.

Regarding diffuse versus specular content, LFP uses the same fixed parameter budget per primitive, including its three directional ASG lobes (Eq. 8). Specular appearance is modelled by the directional lobes of existing primitives rather than by extra material-specific parameters. The total primitive count depends on scene complexity; Tab. 15 gives it per category, including the refractive scenes. The tested refractive scenes have a similar average primitive count to the Tanks and Temples scenes (1.01 M vs. 1.00 M); LFP does not show a substantially higher primitive count on these refractive scenes.

Q2. Poor coverage: capture or charts?

Our experiments indicate that failures under poor coverage stem mainly from insufficient capture rather than from the charts: content not observed by any training ray cannot be inferred from the images. How the charts are constructed has little effect. Placing the 16 chart normals by K-means or on a uniform Fibonacci lattice gives 28.02 / 0.820 / 0.247 vs. 28.02 / 0.818 / 0.245 (PSNR / SSIM / LPIPS) on Mip-NeRF 360 (Tab. 13), and once several charts are used, their number changes little (Tab. 8, PSNR / SSIM / LPIPS). We thus can tell from the experiment results that charts are needed so that every captured ray has well-conditioned coordinates (a single chart is not enough, Fig. 3 (b)), but beyond that their arrangement is not the limiting factor.

Q3. Sparse views (new Tab. 14)

We evaluate sparse-view performance on LLFF with three training views selected as in FSGS. Both methods use the same training views and the same COLMAP points triangulated from these views.

MethodPSNR ↑SSIM ↑LPIPS ↓
3DGS14.070.3370.495
LFP17.580.5430.332

On average over the eight scenes, LFP outperforms 3DGS by a large margin on all three metrics (+3.5 dB PSNR, Tab. 14), without any sparse-view regularizer. We attribute this in part to the explicit epipolar coupling, which ties rays observing the same content across the available views.

Thank you for the detailed response. The clarifications are helpful and have addressed my main questions and concerns.

Reviewer_N92SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a light field-based alternative to NeRF and GS representations for novel view synthesis (NVS). Because the method directly captures light rays (rather than using an intermediate representation), it has the advantage that it can render at different resolutions (zoom in / out), lens blur, lens distortion, etc.

The method is based on differentiable point-based optimization in the 4D ray space. The scene is modeled as a set of atlases (16 in practice, L254). Each atlas is a pair of parallel planes defining rays that pass through both planes (in the traditional light field sense). All the rays that meet at a certain point on a third parallel plane are grouped together into a ray group. Primitives are associated with each group, and nearby rays are rendered using the primitives of the nearby group.

Unlike NeRF, GS, etc., the proposed method makes no attempt to capture or model the 3D geometry of the surfaces in the scene. Therefore, it is not designed for 3D reconstruction.

Critical strengths
  • This is a novel approach to novel view synthesis (NVS).
  • Light fields are a natural choice for NVS, and this paper extends the classic approach with learned primitives.
  • The method is competitive with existing methods, generally performing 1st or 2nd place.
Critical weaknesses
  • The method's scope is fairly narrow, because it only addresses novel view synthesis (NVS). No 3D structure is captured, so applications for robotics, physical AI, AR/VR/XR, etc. will not benefit.
  • The paper is a bit difficult to read in places, with terms not clearly defined. The math is very hard to follow, even with some pre-existing background knowledge.
  • The core method is not clearly explained.
Questions for authors
  • A better diagram is needed to follow the math. If there is not room in the main paper, put this in supplemental. Eq (1) is mostly understandable, but it would be easier if 'd' were labeled in the figure.
  • Why introduce the notation of '\mu^u_i' etc in Eq (3)? Why not just say '\mu = (u_i, v_i, s_i, t_i)'? Seems unnecessarily complicated. If it is not the case that '\mu^u_i = u_i', please explain.
  • Eq (3) should not use braces {}, which indicates an unordered set. Use brackets or parentheses instead. Similarly, don't use calligraphic for 'P', which makes it look like a set. Wait, now that I'm re-reading it, the problem is where you say, right before Eq (3), "Primitive i is specified by". This is not true, as the equation actually describes the set of primitives, not a single primitive.
  • L140 the relationship between disparity 'p_i' and the offset 'z' is not clear to me. First of all, since 'z = 1 / p_i', why don't you use the subscript 'i' for 'z' too? As in, 'z_i'? Also, should 'z' be considered an offset, or something else? It would be good to give it a name. L142 the rearrangement is probably straightforward, but I can't do it in my head to verify. This is where a figure could help, b/c I'm having a hard time visualizing what "disparity" means in this context. I'm assuming that it's similar to binocular stereo disparity, but it's not obvious to me.
  • Eq (4) looks like it should be easy to derive, but I'm struggling with it. Does it arise directly from Eq (2)? This is where the 'mu' notation definitely is making it more complicated.
  • L161 should say something about RGB or color, for clarity.
  • L122 Rotation matrix R should be in SO(3), not R^{3x3}.
  • L155 "is the coordinate" > "are the coordinates"
  • L182 partition by disparity. What if disparity varies continuously, e.g., a slanted plane?
  • L211 "with equality when the query lies in the primitive’s ray group". What does this mean? What does it mean for a query to lie in a group? And why would this lead to equality? From looking at the equation, it appears that equality would happen when we're at the center of the Gaussian ('mu'), so that G=1.
  • L215-221. This description requires a diagram to understand it. Without a diagram, I cannot follow it.
  • L218 disparity does not define a single depth -- why not?
  • L259. "a tangential coordinate on that plane". Do you mean "a pair of tangential coordinates on that plane"? Otherwise, it makes no sense.
  • L261. The method is initialized using structure from motion (SfM). What if that fails?
  • The term "atlas" is not standard (as far as I can tell), and it is not clearly defined. The term is first used (L253) without a definition, so the reader is left to assume that it must refer to a pair of parallel planes. I would suggest using a standard term (if such a term exists) or clearly define the new term.
  • "Transparent object reconstruction" does not refer to 3D reconstruction. Perhaps a different term should be used to avoid confusion.
7 comments
AuthorsOpenReview ↗

Response to Reviewer N92S — Part 1/2

We thank the reviewer for the careful reading. We have adopted all notation comments; they are marked in blue in the revised PDF. We first address the scope concern.

"No 3D structure is captured"

Scope. Like standard NeRF and 3DGS, LFP is designed for novel view synthesis rather than surface reconstruction. Standard 3DGS, for example, is primarily optimized with photometric supervision, whereas dedicated variants such as 2DGS and GOF add geometric constraints for surface reconstruction. Our claims concern novel view synthesis and imaging effects that can be modelled directly in ray space. Accordingly, we compare LFP with NVS methods.

Geometric interpretation and depth rendering. LFP is parameterized entirely in 4D ray space, without explicitly representing 3D primitives or surface geometry. The disparity parameter encodes the geometric relationship among the rays in each group, from which a common 3D location can be derived as (Xi=ξie1+ηie2+n/piX_i = \xi_i e_1 + \eta_i e_2 + n/p_i) (Sec. 2). This derived location is not an explicit scene representation, but can serve as an auxiliary depth reference when compositing primitives across charts (Sec. 3.1, Fig. 3(c)). Furthermore, depth can be rendered using the same compositing weights as colour, without requiring a separate 3D geometry representation.

Quantitative evidence (new Tab. 10, per scene in Tab. 11, visualizations in Fig. 12, all in Sec. F). Following the geometry evaluation protocol of 2DGS (Huang et al. 2024) and GOF (Yu et al. 2024), we compare LFP and 3DGS on the six scenes in Tanks and Temples: Barn, Caterpillar, Courthouse, Ignatius, Meetingroom and Truck, with the official trajectory alignment, which is independent of either method. The results are as follows:

MethodAbsRel ↓δ<1.01 ↑δ<1.05 ↑δ<1.25 ↑F@1cm ↑F@2cm ↑
3DGS0.2019.0%32.3%74.7%0.2490.443
LFP0.17220.3%60.4%87.8%0.2800.476

Depth is the alpha-normalized composited depth on 45 evenly spaced views per scene, scored on the same pixels for both methods. The F-scores evaluate the common locations XiX_i of LFP's ray groups and the 3DGS Gaussian centres, with the same opacity threshold of 0.1 and the official crop. With rendered depth d and laser depth d∗d_*, AbsRel is the mean of ∣d−d∗∣/d∗|d − d_*|/d_*, and δ<τδ<τ is the fraction of pixels with max(d/d∗d/d_*, d∗/dd_*/d) < ττ; the fraction within about 5% of the laser depth (δ<1.05) nearly doubles (60.4% vs. 32.3%). F@1cm / F@2cm is the F-score (harmonic mean of precision and recall) between these locations and the scan at a distance threshold of 1 cm / 2 cm (Sec. F). These results show that the depth and the ray-group locations implied by LFP's ray-space parameters are geometrically meaningful, without any explicit 3D geometry in the representation.

Clarity: new diagrams

  • Fig. 2 (revised). (a) labels the ray direction d. (b) illustrates the ray group of primitive i: all its rays pass through (ξi,ηiξ_i, η_i) on the plane with offset 1/pi1/p_i. (c) uses the anchor (ui,vi)(u_i, v_i), (si,ti)(s_i, t_i) and draws the residual ri(x)/pir_i(x)/p_i.
  • Fig. 3 (new, Sec. 3.1). It illustrates the compositing order. (a) Within one chart the primitives' planes are parallel, so every query meets them in the order of their disparities. (b) A query nearly parallel to a chart's planes has diverging coordinates in that chart; another chart with a different normal gives the same query finite coordinates. (c) Across charts, primitives are ordered by the camera-space depths of their points XiX_i and XjX_j. This answers L215–221 and L218 (in the roriginal manuscript).
  • Fig. 6 (new, Sec. B). (a) Side view: a shift ΔΔ of the uvuv crossing moves the stst crossing by (1−pi)Δ(1−p_i)Δ, giving the coupling s=(1−pi)u+piξis = (1−p_i)u + p_i ξ_i. (b) Epipolar-plane image of the same rays, with equal scales on both axes. The ray group forms a line of slope 1−pi1 − p_i, and the residual ri=pi(ξ−ξi)r_i = p_i(ξ − ξ_i) is the query's displacement from this line along the ss axis.
AuthorsOpenReview ↗

Response to Reviewer N92S — Part 2/2

  • Eq. (1), label dd. dd is now labelled in Fig. 2 (a).
  • μiuμ^u_i notation. μiuμ^u_i was simply the uu-coordinate of the anchor ray. We now write μi=(ui,vi,si,ti)μ_i = (u_i, v_i, s_i, t_i) throughout the main text and the appendix. The convention is stated explicitly in Sec. 2: a subscript ii marks a quantity stored by primitive ii, while unsubscripted coordinates belong to the ray being evaluated.
  • Eq. (3). Eq. (3) describes one primitive. A primitive is now denoted Pi=(μi,pi,Σi,oi,ci)P_i = (μ_i, p_i, Σ_i, o_i, c_i) (Sec. 2).
  • L140, zz vs. pip_i. Here zz is a general plane offset: Eq. (2) holds for any plane parallel to the two parameterization planes, and z=0z = 0 and z=1z = 1 give the uvuv and stst planes themselves. We keep zz unsubscripted and now name it explicitly (the offset of the plane along nn). A primitive's plane is referred to as the plane with offset 1/pi1/p_i, so no ziz_i is needed. Disparity is the inverse depth along nn, measured in units of the gap between the two reference planes ([1, 2]); the content lies on the plane z=1/pz = 1/p. A viewpoint shift ΔΔ of the uvuv crossing moves the stst crossing by (1−p)Δ(1 − p)Δ, i.e., it induces a relative image displacement of −pΔ−pΔ. This is the usual stereo interpretation of pp as relative image displacement per unit viewpoint motion (Sec. B, Fig. 6), and it explains the EPI slope 1−p1 − p. The term is introduced in Sec. 2 and derived in detail in Sec. B.
  • L142, the rearrangement. We set z=1/pz = 1/p in Eq. (2) and multiply by p:p(ξ,η)=(s,t)−(1−p)(u,v)p: p(ξ, η) = (s,t) − (1−p)(u,v). This step is now stated in Sec. 2.
  • Eq. (4). Yes, directly. Evaluating Eq. (2) at z=1/piz = 1/p_i for the query xx and for the anchor μiμ_i gives pi(ξ,η)=(s,t)−(1−pi)(u,v)p_i(ξ,η) = (s,t) − (1−p_i)(u,v) and pi(ξi,ηi)=(si,ti)−(1−pi)(ui,vi)p_i(ξ_i,η_i) = (s_i,t_i) − (1−p_i)(u_i,v_i), and subtracting the two gives the second expression in Eq. (4). This is now stated at Eq. (4).
  • L161. cic_i is the RGB base colour; this is now stated. L122. R∈SO(3)R ∈ SO(3). L155. "are the coordinates". L259. "a pair of tangential coordinates". All four are fixed.
  • L182, continuously varying disparity. Disparity is not quantized. Every primitive has its own continuous, learned pip_i. Here the partition is defined separately for each primitive by its disparity pip_i. It is used only to measure the distance between a query and that primitive's ray group. A slanted surface is covered by many primitives whose pip_i vary continuously across it.
  • L211, "the query lies in the ray group". The Gaussian is over the residual ri(x)r_i(x), not over x−μix − μ_i. A query lies in the group when it satisfies the group's coupling, that is, when it passes through (ξi,ηi)(ξ_i, η_i) on the plane with offset 1/pi1/p_i. Then ri=0r_i = 0, Gi=1G_i = 1 and αi=oiα_i = o_i. This holds for every ray of the group, not only for the anchor (Sec. 3.1; Fig. 6 (b), where the whole line has ri=0r_i = 0).
  • L215–221 and L218. We rewrote this passage (Sec. 3.1) and illustrate it with the new Fig. 3 (a)–(c). Within a chart, all primitive planes are parallel. Their intersections with any query are therefore ordered by disparity, so the primitives of a chart are sorted by pip_i once per view. Across charts, disparities are measured along different normals and are not directly comparable. Moreover, a tilted primitive plane has no single camera depth. We therefore sort primitives of different charts by the camera-space depth of the point XiX_i, where the primitive's response peaks.
  • "Atlas". We dropped the term. Each plane pair is now called a chart, defined at first use (Sec. 2) as one light-field coordinate system: a plane pair with its normal n and tangent basis (e1,e2)(e₁, e₂). A scene uses 16 charts (Sec. 3.2; Fig. 3 (b)), and the term is used consistently in the text, captions, tables and appendix.
  • "Transparent object reconstruction". This is renamed refractive novel view synthesis throughout.
  • L261, if SfM fails. SfM points are only starting values and are not part of the representation (Sec. 3.2). We trained both methods from the random initialization of the 3DGS paper: 100k points drawn uniformly in a cube three times the camera bounding box, the same points for both methods, on Mip-NeRF 360 (new Tab. 12, Sec. F):
MethodSfM init PSNRRandom init PSNR / SSIM / LPIPS
3DGS27.4525.54 / 0.752 / 0.332
LFP28.0226.51 / 0.769 / 0.288

LFP loses less than 3DGS without SfM (−1.51 vs. −1.91 dB; SfM values as in Tab. 1). Even with random initialization, LFP outperforms 3DGS by about 1 dB. Camera poses are required, as for all compared methods.


[1] Jin-Xiang Chai, Xin Tong, Shing-Chow Chan, and Heung-Yeung Shum. Plenoptic sampling. In SIGGRAPH, 2000.

[2] Sven Wanner and Bastian Goldluecke. Globally consistent depth labeling of 4D light fields. In CVPR, 2012.

Thanks to the authors for their careful response. It is fair for 3D structure to be outside the scope of the work, and the fact that the proposed LFP performs better than 3DGS in this regard is therefore even more impressive.

I also appreciate the updates to Fig 2 and the new Fig 3, which are both helpful in understanding the method. My only remaining concern with Fig 2 is that the origin of the coordinate system is still not clear. As a result, some of the vectors and equations remain confusing (more details below).

I also appreciate the new Fig 6. Several questions / comments:

  • Fig 6 a). Since pi=1/zp_i = 1/z, and z=1z=1 for the stst plane (L144-145), then pi=1/1=1p_i=1/1=1, and therefore the vertical quantity labeled as (1−pi)Δ(1-p_i)\Delta on the stst plane (L813) should equal (1−1)Δ=0(1-1)\Delta = 0. But clearly this quantity is not equal to zero, as seen in the figure.
  • Fig 6 a). Similarly, z=0z=0 for the uvuv plane (L144-145), so pi=1/0=∞p_i=1/0=\infty. Therefore, the same vertical quantity should collapse to Δ\Delta for the uvuv plane (L813), but instead it collapses to (1−pi)Δ=(1−∞)Δ=−∞(1-p_i)\Delta = (1-\infty)\Delta = -\infty.
  • Fig 6 a). L830. It would be helpful to draw rir_i on the diagram somewhere. As far as I can tell, rir_i (that is, the component of rir_i in the plane) is equal to tan⁡θ\tan \theta, where θ\theta is the angle between the rays.
  • Fig 6 b). From the eqn relating ss and uu (L829), the yy-intercept should be piξip_i\xi_i. But the line in the diagram goes through the origin (y-intercept = 0).
  • Fig 6 b). L831. If the purpose of the diagram is to provide intuition behind the math, then it fails here. It is not clear how Eq. (4) which defines rir_i in terms of s,si,u,ui,s, s_i, u, u_i, and pip_i relates to the geometric diagram.

Please explain these discrepancies, and whether rir_i can be illustrated in the diagram.

Reviewer_yaAMOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents light field primitives that represent a scene as a compact set of differentiable primitives in the classical two-plane parameterization. Each primitive holds one learned record for a group of rays and answers a query by how far the query departs from the coupling that defines the group, without projecting or marching rays. A scene is optimized from posed images and rendered in real time, as with point-based methods, using gradient descent. The authors evaluate their method on established NVS benchmarks and confirm its effectiveness with solid results across all tests.

Critical strengths

The paper makes several solid technical contributions and is exhaustive in its evaluation:

Specifically, the authors exploit the affine coupling that defines a ray group both to gather the rays sharing a primitive’s record and to grade a query against them, giving a response function that can render in real time within the 4D ray space.

This is adequately derived and analyzed in the paper. The contribution is effective and confirmed to be functional with large margins across common NVS benchmarks. The authors also show that the method can be extended to multi-scale anti-aliasing, defocus deblurring with post-hoc refocusing, fisheye rendering, and transparent-object reconstruction, at a quality comparable to frameworks specialized for each effect.

Overall, the experimental findings consistently show a margin compared to existing methods in almost all qualiative and quantiativel evaluations presented in the paper.

Critical weaknesses

No major critical weaknesses. The paper comes with an extensive appendix that answers my questions.

Questions for authors

Do the authors plan to release code to better ensure reproducibility?

1 comment
AuthorsOpenReview ↗

Response to Reviewer yaAM

We thank the reviewer for the strong support. Yes. Upon acceptance we will release the complete codebase, including the CUDA renderer and the training and evaluation scripts for all five experimental setups, together with the trained checkpoints.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comments

We thank the ACs and all reviewers for their invaluable time and comments. We have uploaded a revised PDF in which every change is marked in blue. The main changes are:

  • Clarity (RN92S). Fig. 2 is redrawn by labeling the ray direction d. A new Fig. 3 (Sec. 3.1) shows how primitives are ordered within one chart and across charts, and why several charts are needed. A new Fig. 6 (Sec. B) shows disparity, the coupling and the residual in a side view and in the epipolar-plane image. The notation is simplified to μ_i = (u_i, v_i, s_i, t_i), key derivation steps are now given in the main text, and the non-standard term "atlas" is replaced by "chart", defined at first use.
  • 3D structure, random initialization, sparse views (RN92S, R6Z2h). These are in the new Sec. F: Tab. 10–11 and Fig. 12 (3D structure), Tab. 12 (random initialization), Tab. 14 (sparse view), with the common protocol.
  • Neural light fields (R6Z2h). A detailed paragraph is added to the related work (Sec. A).
  • Primitive budget by category (R6Z2h) is in Tab. 15, and chart (atlas) construction are in Tab. 13.

For all experiments, the compared methods use the same input images, train/test split, initialization, and evaluation protocol. We use the official implementations of other methods without modification and LFP in its main-table configuration without per-experiment tuning.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.