Pith. sign in

REVIEW 4 major objections 5 minor 52 references

The paper claims that conditioning a neural acoustic field on explicit local geometry—ray-hit distances, surface normals, and distance statistics probed from a rough room mesh—yields more accurate room impulse response predictions than imag

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 16:12 UTC pith:TVTICSNR

load-bearing objection MiNAF's ray-mesh feature set is a genuinely new idea, but the central causal claim is undermined by a Schroeder loss that directly optimizes the evaluation metrics and by an ablation table that contradicts the paper's own narrative. the 4 major comments →

arxiv 2509.15210 v2 pith:TVTICSNR submitted 2025-09-18 cs.SD cs.AIcs.LG

Explicit Context-Driven Neural Acoustic Modeling for High-Fidelity RIR Generation

Chen Si , Qianyi Wu , Chaitanya Amballa , Romit Roy Choudhury This is my paper
classification cs.SD cs.AIcs.LG
keywords room impulse responseneural acoustic fieldsexplicit geometryray-mesh intersectionbinaural audioreverberation timefew-shot learningindoor sound simulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

MiNAF proposes to condition a neural acoustic field on explicit local geometry: at each transmitter and receiver location it casts many rays into a rough room mesh and reads off first-hit distances, surface normals, neighborhood distance means and standard deviations, and an occupancy histogram. The claim is that these direct geometric features are a better context for predicting a room impulse response than the image-based or learned global contexts used by previous neural implicit RIR models. A simple MLP then predicts the STFT log-magnitude and instantaneous frequency, with a Schroeder-curve loss to sharpen energy decay. On benchmark rooms, the approach reports lower T60 and EDT than compared baselines at comparable C50, including a 22% T60 improvement over the strongest image-based baseline, and it keeps working when training data is cut to 5–10% and when the mesh carries moderate noise. Why care: if this holds, realistic binaural audio for AR/VR could be synthesized from a phone-generated mesh and a few microphone measurements rather than from dense visual context.

Core claim

MiNAF's central discovery is that explicit local geometry—obtained by ray-casting from the query point against a rough mesh and summarizing first-hit distances, normals, and distance statistics—is a sufficient and efficient context for a neural network to synthesize the room impulse response spectrum. Under the same reconstruction method, MiNAF reports a 22% lower T60 than the best image-context neural baseline with nearly identical C50; with ground-truth phase it reports 40% lower T60 and 26% lower C50 than an audio-visual NeRF baseline. The paper interprets this as evidence that direct geometric cues guide the network's magnitude and energy-decay predictions better than global image featur

What carries the argument

The central mechanism is the context-retrieval stage: for each transmitter and receiver point, N=1024 rays from a Fibonacci lattice are intersected with a room mesh; their points of first hit supply distances, surface normals, distances' neighbor mean and standard deviation, and an occupancy histogram. These feature sets are projected into a shared latent space, concatenated with encoded positions, and element-wise multiplied by a positional encoding of time before a five-layer MLP outputs the STFT log-magnitude and instantaneous-frequency columns; a Schroeder-curve loss term aligns the energy-decay envelope.

Load-bearing premise

The premise that carries the argument is that a rough room mesh, and especially the first-hit ray distances and surface normals at the transmitter and receiver, capture enough of the geometry to determine the full impulse response including late reverberation.

What would settle it

Concrete test: train MiNAF with the full feature set, then replace each first-hit distance vector with random distances drawn from the same room's distance histogram while keeping normals and statistics unchanged; if T60 and EDT remain the same, per-ray geometry is not the carrier. Alternatively, a room whose reconstructed mesh is accurate to under 30 cm but where MiNAF's T60 error jumps to baseline level would falsify the robustness claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, room acoustics can be modeled from a rough mesh plus sparse RIR recordings, bypassing image-based context retrieval and its camera-coverage assumptions.
  • With 10% of training data, MiNAF reports better T60 than a full-data image-context baseline; this suggests acoustic-field learning can be data-efficient when geometry is explicitly provided.
  • Moderate mesh noise (up to roughly 30 cm vertex displacement) causes graceful degradation; even a mesh reconstructed from casual multi-view photos keeps performance close to the true mesh, so the method is deployable with real-world reconstruction pipelines.
  • Ablations show that removing surface normals hurts most, indicating that the direction of the last reflection, not just distance, carries acoustic information; removing neighbor distance statistics and the occupancy histogram also degrades T60 and EDT.
  • Predicted phase remains the bottleneck: the predicted-phase variant underperforms ground-truth-phase and random-phase variants on several metrics, and the paper cautions against using Griffin-Lim for real RIRs despite using it for comparison.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: the same ray-mesh feature set could be attached to other neural acoustic field architectures to test whether explicit geometry transfers beyond MiNAF's particular MLP and context-fusion scheme.
  • Editorial extension: because the features are local and mesh-derived, they might generalize to unseen rooms after a cheap per-room calibration step; the paper only evaluates same-scene train/test splits, so cross-scene transfer with few-shot fine-tuning is a natural next test.
  • Editorial extension: the 50 cm noise failure and multi-room outlier rays suggest the method's robustness is bounded; distance-based outlier rejection or multi-scale ray sampling might extend the operating range beyond the reported conditions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes MiNAF, a neural implicit model for room impulse response (RIR) generation that augments the usual transmitter/receiver coordinates with explicit local geometric features extracted by ray-casting against a rough room mesh. These features include first-hit distances, surface normals, proximity statistics, and an occupancy histogram. The model predicts STFT log-magnitude and instantaneous frequency, from which the RIR is reconstructed using one of several phase strategies (ground-truth, predicted, random, or Griffin--Lim). Experiments on the SoundSpaces dataset compare MiNAF against five neural implicit baselines and two codec baselines, reporting lower T60 and EDT, comparable C50, robustness to reduced training data, and tolerance to noisy or reconstructed meshes. The central claim is that explicit local geometry yields more accurate RIR predictions than image-based or global-context approaches.

Significance. If established, the claim that direct ray-mesh features improve neural RIR generation would be a useful contribution to audio-visual scene modeling. The paper has strengths: it evaluates on a standard benchmark (SoundSpaces) across multiple room types, considers several phase-reconstruction variants for fair comparison, and includes robustness studies with Gaussian mesh noise and a VGGT reconstruction. However, the causal attribution of the improvements to explicit geometry is not currently supported because (a) the training loss includes a Schroeder-curve term that directly optimizes the headline decay metrics (T60/EDT), which most baselines do not use, and (b) the ablation evidence is internally inconsistent, with one variant improving when context is removed. The single-run evaluation further weakens confidence in small metric differences. The idea is promising, but the evidence needs reworking to isolate the effect of the geometric context.

major comments (4)
  1. [Model Training] The loss is defined as L = L1(spectrum) + α × L_Schroeder(waveform). T60 and EDT are derived from the Schroeder decay curve, so this loss directly optimizes the headline metrics. The five neural baselines (NAF, NeRAF, etc.) are not described as using such a loss. Consequently, the 22% T60 improvement of MiNAF(GLim) over NeRAF in Table 1 may be due to the loss term rather than the proposed explicit geometric features. To support the causal claim, the authors should retrain MiNAF with α=0 (or with a pure spectral loss) and compare; ideally they should also retrain or re-evaluate baselines with the same Schroeder loss. Without this control, the central attribution is confounded.
  2. [Table 3] The ablation table contradicts the paper's assertion that removing context degrades performance. For MiNAF(GLim), removing the entire context ('w/o C') improves T60 from 1.40 to 1.27 and C50 from 0.4038 to 0.3847 (lower is better for both). This is the opposite of what the text claims in the 'Context Information' paragraph. While MiNAF(PreP) and MiNAF(RanP) show degradation when context is removed, the GLim variant does not. A coherent ablation should either show consistent degradation across all reconstruction methods or explain why the effect differs. Because results are single-run (Appendix B), this could be noise; multi-seed ablations are needed.
  3. [Appendix B (Experimental Runs)] The paper states that every reported result, including ablations, comes from a single training run. Many of the claimed differences are small (e.g., MiNAF(GTP) T60 1.49 vs MiNAF(RanP) 1.42; some Table 3 differences are ~0.1). Without error bars or multiple seeds, it is impossible to determine whether the observed improvements are statistically significant. At minimum, the key comparisons (Table 1, Table 3, and Fig. 5) should be accompanied by variances over at least 3–5 seeds.
  4. [Experiments (Baselines)] It is unclear whether the baseline numbers in Table 1 are from prior publications or were retrained/re-evaluated under identical conditions (same data split, same preprocessing, same phase-reconstruction algorithm). The superscript notation indicates matching phase reconstruction methods, but the text does not state that baselines were rerun. If the numbers are taken from original papers, differences in training protocol, data splits, and evaluation details may bias the comparison. The authors should clarify and, if necessary, rerun all baselines under the same pipeline.
minor comments (5)
  1. [Model Training] The Schroeder loss term is not fully specified. Please provide the formula for L_Schroeder and the value of α, as it is essential for reproducibility and for understanding the loss balance.
  2. [Table 3] The notation '- w/on' is ambiguous; it should be written '- w/o n' (without normals). Also, the caption says 'Top-1, top-2, and top-3 results are highlighted' but the table appears to bold only cell values, which is confusing.
  3. [Appendix C] The text admits that the energy-curve loss may cause overfitting to T60 in room 2. This admission underscores the confound raised in the major comments and should be addressed in the main text rather than only in the appendix.
  4. [Figure 5] The subplot labels ('T60 (%)', 'C50 (dB)', 'EDT (sec)') have different scales and units. Adding error bars or at least markers for each data point would improve readability. Currently the curves are hard to resolve at low data percentages.
  5. [References] Several references lack full publication details (e.g., Li et al. 2025, Wang et al. 2024). Please complete the bibliography entries.

Circularity Check

0 steps flagged

No significant circularity: MiNAF's claims are evaluated against external SoundSpaces baselines; the Schroeder-loss metric alignment is a confound, not a circular reduction.

full rationale

I walked the derivation chain. The geometry context is defined via explicit ray-mesh queries (distances, normals, proximity statistics, occupancy), concatenated with encoded positions, and fed to an MLP to predict STFT spectra. This is a standard supervised learning setup with held-out test positions; the prediction is not defined in terms of the target and no equation reduces RIR to the input context by construction. The paper contains no self-citations (the author list has no overlap with the cited prior work) and invokes no uniqueness theorems or ansatz smuggled through self-citation. The only notable concern is that the training loss includes L_Schroeder(waveform), which directly targets the energy-decay curve from which T60 and EDT are computed; this could confound comparisons with baselines that may not use such a loss, and Table 3 shows the GLim ablation 'w/o C' actually improving T60 and C50, while Appendix B reports that each result comes from a single training run. These are correctness/robustness issues, not circularity: the evaluation is against external benchmarks, and the geometry ablations use the same loss, so the central geometric-context claim is not statistically forced by construction. Score 1 reflects the mild training-to-metric alignment rather than circular reasoning.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

This is a supervised learning method, so the 'derivation' is an empirical fit. The free parameters are architecture/data hyperparameters, not physics constants. The axioms are the modelling choices: mesh availability and the sufficiency of local geometric statistics to predict global reverberation.

free parameters (6)
  • Loss balancing factor alpha
    Scales the Schroeder curve loss relative to L1 spectrum loss; value not reported, tuned by experiment, and directly impacts decay metrics (T60/EDT).
  • Context latent dimension h
    Dimension of the projected feature vectors C_p; not specified numerically in text, affects capacity.
  • Number of rays N = 1024
    Hand-chosen; ablation shows performance degrades with fewer rays (Fig. 8).
  • Nearest neighbors for proximity statistics N_theta = 8
    Hand-chosen neighbor count for mean/std of distances.
  • Histogram thresholds (N_tau=8, delta from 0.5m to 10m) = 8 thresholds, linearly spaced 0.5-10m
    Hand-chosen occupancy bins; scale depends on room size.
  • Positional encoding bands L = 10
    Number of frequency bands for sinusoidal encodings of positions and time.
axioms (5)
  • domain assumption A rough 3D mesh of the room is available as input and approximates the true acoustic geometry.
    The whole context extraction queries the mesh with rays (Geometry Context Collection). The paper tests noise robustness, but the assumption is central.
  • domain assumption First-hit ray distances, surface normals at the hits, and local distance statistics at the transmitter/receiver positions suffice to determine the RIR.
    The feature set in Geometry Context Collection is asserted to capture local geometry; no proof that these statistics contain enough information for late reverberation.
  • domain assumption RIRs can be faithfully predicted from log-magnitude and instantaneous frequency spectra via STFT and inverse STFT.
    Used in Task Definition; IF avoids phase wrapping, but phase reconstruction remains a known weak point (Appendix D, E).
  • domain assumption SoundSpaces simulated RIRs are an adequate ground truth for training and evaluation.
    All experiments use SoundSpaces (Chen et al. 2020), which are simulated, not measured, RIRs.
  • standard math A 5-layer fully-connected MLP with the described input features is capable of representing the acoustic field.
    Relies on universal approximation; not formally guaranteed for this problem.

pith-pipeline@v1.3.0-alltime-deepseek · 16691 in / 13278 out tokens · 128323 ms · 2026-08-04T16:12:51.652499+00:00 · methodology

0 comments
read the original abstract

Realistic sound simulation plays a critical role in many applications. A key element in sound simulation is the room impulse response (RIR), which characterizes how sound propagates within a given space. Recent studies have applied neural implicit methods to learn RIR using context information collected from the environment, such as scene images. However, these approaches do not effectively leverage explicit geometric information from the environment. To further exploit neural implicit models with direct geometric features, we present MiNAF, which queries a rough room mesh at given locations and extracts distance distributions as an explicit representation of local context. Our approach demonstrates that incorporating explicit local geometric features can better guide the model in generating more accurate RIR predictions. Through comparisons with conventional and state-of-the-art methods, we show that MiNAF performs competitively across various evaluation metrics.

Figures

Figures reproduced from arXiv: 2509.15210 by Chaitanya Amballa, Chen Si, Qianyi Wu, Romit Roy Choudhury.

Figure 1
Figure 1. Figure 1: Task Overview. To record audio, a loudspeaker (T x) and several microphones (Rx) in red are placed at dif￾ferent known locations within the room. The room’s rough layout is available from images, 3D mesh, or LiDAR scans. MiNAF learns to reconstruct the Room Impulse Response (RIR) from sparse audio recordings at red microphones and explicit geometric features from probing the environment. The model can then… view at source ↗
Figure 2
Figure 2. Figure 2: Workflow Overview. A context retriever first extracts physical features at the transmitter and receiver locations, generating context vectors CTx and CRx. These are concatenated with their respective positions to form a comprehensive representation of the environment. Next, positional encoding is applied to the time t, channel c, and orientation θ. The encoded time is then element-wise embedded into the co… view at source ↗
Figure 3
Figure 3. Figure 3: Demonstration of Context Collection. For each transmitter and receiver location p, we uniformly sample N rays with Fibonacci’s lattice. Then, we measure the distance for a ray from the center p to its point of first hit (PoFH). A general description of the distribution of neighboring rays’ distances is also included in this measurement. mesh, we derive a feature vector characterizing the spatial properties… view at source ↗
Figure 4
Figure 4. Figure 4: Comparisons: (1) Average SNR and (2) Average [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Few-shot Experiment. MiNAF trained with 10% of data outperforms AV-NeRF trained with the entire dataset. MiNAF also outperforms NeRAF in its ability to quickly adapt to few-shot datasets. phases. In terms of PSNR, MiNAF produces better results than INRAS on synthesized RIRs overall methods. Ablation Studies Following Brunetto et al. (2024), we experiment with room 2 to evaluate the effectiveness of each mo… view at source ↗
Figure 8
Figure 8. Figure 8: Probe Density Experiment: Performance of MiNAF tends to decrease with fewer rays, etc. ports the effectiveness of fusing time embeddings into con￾text features in enabling the model to capture the scene’s spatiotemporal characteristics more accurately, reinforcing their importance in improving RIR prediction quality. Probe Density. We evaluate the effect of varying the num￾ber of emanating rays sampled at … view at source ↗
Figure 9
Figure 9. Figure 9: Log-magnitude and IF Spectra. Here we show the predicted and ground-truth log-magnitude and instantaneous frequency (IF) spectra of samples from four representative rooms: room 2, office 3, frl apartment 2, and apartment 1. The visualizations demonstrate that MiNAF excels in capturing the magnitude spectrum with high fidelity. However, learning phase information proves to be more challenging. While the pre… view at source ↗
Figure 10
Figure 10. Figure 10: RIR in Time Domain and its Energy Curve. We show the RIR in the time domain and its energy decay for a random sample in Room 2. In each subplot, the dashed line shows the Schroeder Curve, which represents the RIR’s energy decay. The red, blue, and green markers represent the point of maximum RIR energy, the start point to measure T60 (when energy decays for 5 dB), and the endpoint to measure T60 (when ene… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 6 linked inside Pith

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    B.; and Berkley, D

    Allen, J. B.; and Berkley, D. A. 1979. Image method for efficiently simulating small-room acoustics. The Journal of the Acoustical Society of America, 65(4): 943--950

    1979

  4. [4]

    A.; and Van Waterschoot, T

    Antonello, N.; De Sena, E.; Moonen, M.; Naylor, P. A.; and Van Waterschoot, T. 2017. Room impulse response interpolation using a sparse spatio-temporal representation of the sound field. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 25(10): 1929--1941

    2017

  5. [5]

    Bai, X.; Zhou, J.; Ning, X.; and Wang, C. 2022. 3D data computation and visualization

    2022

  6. [6]

    T.; Mildenhall, B.; Tancik, M.; Hedman, P.; Martin-Brualla, R.; and Srinivasan, P

    Barron, J. T.; Mildenhall, B.; Tancik, M.; Hedman, P.; Martin-Brualla, R.; and Srinivasan, P. P. 2021. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, 5855--5864

    2021

  7. [7]

    Brunetto, A.; Hornauer, S.; and Moutarde, F. 2024. NeRAF: 3D Scene Infused Neural Radiance and Acoustic Fields. CoRR, abs/2405.18213

    arXiv 2024

  8. [8]

    Chen, C.; Jain, U.; Schissler, C.; Gari, S. V. A.; Al-Halah, Z.; Ithapu, V. K.; Robinson, P.; and Grauman, K. 2020. Soundspaces: Audio-visual navigation in 3d environments. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part VI 16, 17--36. Springer

    2020

  9. [9]

    Chen, D.; Li, H.; Ye, W.; Wang, Y.; Xie, W.; Zhai, S.; Wang, N.; Liu, H.; Bao, H.; and Zhang, G. 2024. PGSR: Planar-based Gaussian Splatting for Efficient and High-Fidelity Surface Reconstruction. arXiv preprint arXiv:2406.06521

    Pith/arXiv arXiv 2024

  10. [10]

    Dufter, P.; Schmitt, M.; and Sch \"u tze, H. 2022. Position information in transformers: An overview. Computational Linguistics, 48(3): 733--763

    2022

  11. [11]

    K.; Chen, S.; Gulrajani, I.; Donahue, C.; and Roberts, A

    Engel, J.; Agrawal, K. K.; Chen, S.; Gulrajani, I.; Donahue, C.; and Roberts, A. 2019. Gansynth: Adversarial neural audio synthesis. arXiv preprint arXiv:1902.08710

    Pith/arXiv arXiv 2019

  12. [12]

    E.; Pingali, G.; Min, P.; and Ngan, A

    Funkhouser, T.; Tsingos, N.; Carlbom, I.; Elko, G.; Sondhi, M.; West, J. E.; Pingali, G.; Min, P.; and Ngan, A. 2004. A beam tracing method for interactive architectural acoustics. The Journal of the acoustical society of America, 115(2): 739--756

    2004

  13. [13]

    J.; Kowalski, M.; Johnson, M.; Shotton, J.; and Valentin, J

    Garbin, S. J.; Kowalski, M.; Johnson, M.; Shotton, J.; and Valentin, J. 2021. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF international conference on computer vision, 14346--14355

    2021

  14. [14]

    Griffin, D.; and Lim, J. 1984. Signal estimation from modified short-time Fourier transform. IEEE Transactions on acoustics, speech, and signal processing, 32(2): 236--243

    1984

  15. [15]

    A.; and Duraiswami, R

    Gumerov, N. A.; and Duraiswami, R. 2009. A broadband fast multipole accelerated boundary element method for the three dimensional Helmholtz equation. The Journal of the Acoustical Society of America, 125(1): 191--205

    2009

  16. [16]

    He, Y.; Cherian, A.; Wichern, G.; and Markham, A. 2024. Deep Neural Room Acoustics Primitive. In Salakhutdinov, R.; Kolter, Z.; Heller, K.; Weller, A.; Oliver, N.; Scarlett, J.; and Berkenkamp, F., eds., Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, 17842--17857. PMLR

    2024

  17. [17]

    Huang, X.; Li, H.; Yang, Z.; Wang, Z.; and Liang, X. 2024. 3D visibility-aware generalizable neural radiance fields for interacting hands. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 2400--2408

    2024

  18. [18]

    International Organization for Standardization . 2006. Advanced Audio Coding (AAC). ISO/IEC 13818-7:2006

    2006

  19. [19]

    Jin, W.; and Kleijn, W. B. 2015. Theory and design of multizone soundfield reproduction using sparse methods. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 23(12): 2343--2355

    2015

  20. [20]

    Kingma, D. P. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    Pith/arXiv arXiv 2014

  21. [21]

    Kurucz, P.; Baghaei, N.; Serafin, S.; and Klein, E. 2023. Enhancing Auditory Immersion in Interactive Virtual Reality Environments. In 2023 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct), 789--792. IEEE

    2023

  22. [22]

    Lee, S.; Choi, J.; Kim, S.; Kim, I.-J.; and Cho, J. 2024. Few-shot neural radiance fields under unconstrained illumination. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 2938--2946

    2024

  23. [23]

    Li, L.; Zhang, L.; Wang, Z.; Zhang, F.; Li, Z.; and Shen, Y. 2025. Representing Sounds as Neural Amplitude Fields: A Benchmark of Coordinate-MLPs and a Fourier Kolmogorov-Arnold Framework. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 24458--24466. AAAI Press. Available at: AAAI 2025

    2025

  24. [24]

    H.; Unberath, M.; Liu, M.-Y.; and Lin, C.-H

    Li, Z.; M \"u ller, T.; Evans, A.; Taylor, R. H.; Unberath, M.; Liu, M.-Y.; and Lin, C.-H. 2023. Neuralangelo: High-fidelity neural surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8456--8465

    2023

  25. [25]

    Liang, S.; Huang, C.; Tian, Y.; Kumar, A.; and Xu, C. 2023 a . Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. Advances in Neural Information Processing Systems, 36: 37472--37490

    2023

  26. [26]

    Liang, S.; Huang, C.; Tian, Y.; Kumar, A.; and Xu, C. 2023 b . Neural acoustic context field: Rendering realistic room impulse response with neural fields. arXiv preprint arXiv:2309.15977

    Pith/arXiv arXiv 2023

  27. [27]

    Luo, A.; Du, Y.; Tarr, M.; Tenenbaum, J.; Torralba, A.; and Gan, C. 2022. Learning neural acoustic fields. Advances in Neural Information Processing Systems, 35: 3165--3177

    2022

  28. [28]

    Majumder, S.; Chen, C.; Al-Halah, Z.; and Grauman, K. 2022. Few-shot audio-visual learning of environment acoustics. Advances in Neural Information Processing Systems, 35: 2522--2536

    2022

  29. [29]

    P.; Tancik, M.; Barron, J

    Mildenhall, B.; Srinivasan, P. P.; Tancik, M.; Barron, J. T.; Ramamoorthi, R.; and Ng, R. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1): 99--106

    2021

  30. [30]

    Ni, Z.; Yang, P.; Yang, W.; Ma, L.; and Kwong, S. 2024. ColNeRF: Collaboration for Generalizable Sparse Input Neural Radiance Field. 38(5): 4325--4333

    2024

  31. [31]

    T.; Mildenhall, B.; Sajjadi, M

    Niemeyer, M.; Barron, J. T.; Mildenhall, B.; Sajjadi, M. S.; Geiger, A.; and Radwan, N. 2022. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5480--5490

    2022

  32. [32]

    Ratnarajah, A.; Tang, Z.; and Manocha, D. 2020. IR-GAN: Room impulse response generator for far-field speech recognition. arXiv preprint arXiv:2010.13219

    Pith/arXiv arXiv 2020

  33. [33]

    Reiser, C.; Peng, S.; Liao, Y.; and Geiger, A. 2021. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision, 14335--14345

    2021

  34. [34]

    Richard, A.; Dodds, P.; and Ithapu, V. K. 2022. Deep impulse responses: Estimating and parameterizing filters with deep networks. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 3209--3213. IEEE

    2022

  35. [35]

    Rungta, A.; Schissler, C.; Rewkowski, N.; Mehra, R.; and Manocha, D. 2018. Diffraction kernels for interactive sound propagation in dynamic environments. IEEE transactions on visualization and computer graphics, 24(4): 1613--1622

    2018

  36. [36]

    Savioja, L.; and Valimaki, V. 2003. Interpolated rectangular 3-D digital waveguide mesh algorithms with frequency warping. IEEE transactions on speech and audio processing, 11(6): 783--790

    2003

  37. [37]

    Schissler, C.; Loftin, C.; and Manocha, D. 2017. Acoustic classification and optimization for multi-modal rendering of real-world scenes. IEEE transactions on visualization and computer graphics, 24(3): 1246--1259

    2017

  38. [38]

    Schissler, C.; and Manocha, D. 2016. Interactive sound propagation and rendering for large multi-source scenes. ACM Transactions on Graphics (TOG), 36(4): 1

    2016

  39. [39]

    L.; and Frahm, J.-M

    Sch\" o nberger, J. L.; and Frahm, J.-M. 2016. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR)

    2016

  40. [40]

    L.; Zheng, E.; Pollefeys, M.; and Frahm, J.-M

    Sch\" o nberger, J. L.; Zheng, E.; Pollefeys, M.; and Frahm, J.-M. 2016. Pixelwise View Selection for Unstructured Multi-View Stereo. In European Conference on Computer Vision (ECCV)

    2016

  41. [41]

    Singh, N.; Mentch, J.; Ng, J.; Beveridge, M.; and Drori, I. 2021. Image2reverb: Cross-modal reverb impulse response synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 286--295

    2021

  42. [42]

    J.; Mur-Artal, R.; Ren, C.; Verma, S.; et al

    Straub, J.; Whelan, T.; Ma, L.; Chen, Y.; Wijmans, E.; Green, S.; Engel, J. J.; Mur-Artal, R.; Ren, C.; Verma, S.; et al. 2019. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797

    Pith/arXiv arXiv 2019

  43. [43]

    Su, K.; Chen, M.; and Shlizerman, E. 2022. Inras: Implicit neural representation for audio scenes. Advances in Neural Information Processing Systems, 35: 8144--8158

    2022

  44. [44]

    J.; Li, D.; Langlois, T

    Tang, Z.; Bryan, N. J.; Li, D.; Langlois, T. R.; and Manocha, D. 2020. Scene-aware audio rendering via deep acoustic analysis. IEEE transactions on visualization and computer graphics, 26(5): 1991--2001

    2020

  45. [45]

    Tsingos, N.; Funkhouser, T.; Ngan, A.; and Carlbom, I. 2001. Modeling acoustics in virtual environments using the uniform theory of diffraction. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 545--552

    2001

  46. [46]

    Turki, H.; Ramanan, D.; and Satyanarayanan, M. 2022. Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12922--12931

    2022

  47. [47]

    Wang, J.; Chen, M.; Karaev, N.; Vedaldi, A.; Rupprecht, C.; and Novotny, D. 2025. VGGT: Visual Geometry Grounded Transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2025

  48. [48]

    Wang, S.; Leroy, V.; Cabon, Y.; Chidlovskii, B.; and Revaud, J. 2024. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20697--20709

    2024

  49. [49]

    Xiph.Org Foundation . 2012. Xiph Opus. https://opus-codec.org/

    2012

  50. [50]

    Yu, A.; Ye, V.; Tancik, M.; and Kanazawa, A. 2021. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 4578--4587

    2021

  51. [51]

    Yu, Z.; Sattler, T.; and Geiger, A. 2024. Gaussian Opacity Fields: Efficient Adaptive Surface Reconstruction in Unbounded Scenes. ACM Transactions on Graphics

    2024

  52. [52]

    Zhang, H.; Wang, C.; Tian, S.; Lu, B.; Zhang, L.; Ning, X.; and Bai, X. 2023. Deep learning-based 3D point cloud classification: A systematic survey and outlook. Displays, 79: 102456

    2023