{"id":"beed3fb6-f5a4-4e4a-a4d6-68caf8e21f3b","arxiv_id":"2607.17660","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"RayOcc models each camera ray as a non-normalized Gaussian mixture with Poisson-based occupancy probabilities, allowing multiple depth hypotheses per ray and improving Gaussian-initialized 3D occupancy prediction on nuScenes.","lead":"RayOcc predicts several possible surfaces along each camera ray instead of just one, using a non-normalized Gaussian mixture that turns into occupancy probabilities. On the nuScenes benchmark it reports better 3D occupancy accuracy and cleaner Gaussian initialization than other Gaussian-based methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains are not shown to come from the multi-label Poisson formulation; no ablation replaces it with a normalized single-hypothesis variant, leaving the central causal claim underdetermined.","rationale":"The reader's weakest_assumption identifies a real issue with the point-sampled ray labels, but the more load-bearing gap is the absence of a controlled ablation of the proposed formulation itself. Even with perfect labels, the comparison to VG3T/GaussianFormer-2 cannot separate the multi-label Poisson mechanism from the adaptive sampling, the ray loss, or the increased Gaussian count. A same-pipeline softmax baseline would settle this. The label-sampling issue compounds the problem: because the ray loss uses binary labels derived from a single point per ~0.55m bin, the model receives false negatives on a large fraction of occupied voxels (mean chord length ~0.33m), which could actively suppress the thin-structure predictions that motivate the method. So the two concerns are related, but the missing ablation is more fundamental. The correct verdict remains CONDITIONAL, pending the ablation and release of training details; no change from the reader's verdict.","tokens_in":11549,"tokens_out":12957,"duration_ms":126683,"concrete_test":"Replace the non-normalized intensity + Eq. 6 in RayOcc with a softmax-normalized categorical distribution over the same 128 depth bins (single depth hypothesis per ray), keeping the backbone, mixture network capacity, losses, and Gaussian splatting fixed. Retrain on nuScenes and compare mIoU/IoU to Table I. If the difference is <0.3 mIoU, the multi-label Poisson formulation is not the cause of the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is causal: non-normalized Gaussian mixture intensity converted by Eq. 6 avoids mutual depth competition and therefore yields better occupancy. Yet no experiment isolates this mechanism. Table IV ablates only the activation threshold/sampling, not the intensity-to-probability mapping or the multi-label assumption. VG3T (the closest baseline) differs in depth regression, absence of the auxiliary ray loss, and network details; the 0.78 IoU/0.29 mIoU gain over it is within plausible run-to-run noise given no error bars. The ground-truth ray labels (Eq. 14) are also suspect: one point sample per 0.55 m bin can miss a 0.5 m voxel whose mean chord length is ~0.33 m, so false-negative labels may actively suppress thin structures the method is designed to recover. If lambda_ray is weighted heavily, this biases the mixture hypotheses. Thus the reported SOTA is consistent with the method, but the paper does not demonstrate that the multi-label formulation, rather than the extra Gaussians, the VGGT backbone, or training details, drives the improvement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"RayOcc proposes to model each camera ray's occupancy as a non-normalized Gaussian mixture intensity, converts integrated intensity into interval-wise occupancy probabilities via a Poisson event model (Section III-B, Eqs. 3-6), and uses component amplitudes to adaptively initialize 3D Gaussian primitives for semantic occupancy rasterization. On nuScenes with SurroundOcc annotations, it reports IoU 34.84 and mIoU 22.03, exceeding the VG3T baseline. Ablations show that component-wise sampling improves Gaussian initialization position metrics (Table II) and that the activation threshold has modest impact on occupancy accuracy (Table IV).","tokens_in":11860,"tokens_out":4202,"duration_ms":39457,"significance":"If the causal claim holds, this is a useful conceptual correction: replacing softmax-normalized categorical depth with multi-label intensity aligns ray supervision with the multi-surface nature of volumetric occupancy. The Poisson derivation from intensity to interval existence is clean, and the auxiliary ray loss is well motivated. The use of dense occupancy annotations to generate ray labels is standard supervised training, not circularity. However, the evidence for the central claim is incomplete: no experiment isolates the intensity-to-probability mapping from the rest of the pipeline, and the main comparison with VG3T is close and lacks uncertainty. A carefully controlled ablation and variance reporting would substantially strengthen the paper.","major_comments":[{"comment":"The central claim -- that non-normalized intensity converted by Eq. (6) avoids mutual depth competition and thereby improves occupancy -- is not isolated by any experiment. Table IV varies only the activation threshold and Bernoulli sampling; no variant replaces Eq. (6) with a normalized categorical distribution (e.g., softmax over integrated bin intensities) or removes the auxiliary ray loss while holding the backbone and refinement fixed. Since the compared VG3T baseline differs in depth regression, network details, and training, the reported +0.78 IoU / +0.29 mIoU could plausibly come from the VGGT backbone, the extra Gaussians, or lambda_ray. Please add an ablation that swaps only the ray probability model.","section":"Section III-B / Section IV-E"},{"comment":"Ray labels are generated by sampling a single 3D point at the center of each depth bin. With S=128 bins over [1,71] m, the bin width is ~0.55 m, comparable to the 0.5 m voxel size. A ray can pass through an occupied voxel without containing that voxel's bin-center point, so thin structures and partial voxels will acquire false-negative labels. Since these labels supervise the mixture intensity via L_ray, the bias propagates into the multi-label hypotheses that are the paper's main contribution. Please quantify the label noise, use multiple samples per bin, or derive ray-voxel intersection labels directly.","section":"Section III-E, Eq. (14)"},{"comment":"The abstract and main results claim state-of-the-art performance, but Table I omits several strong baselines that are cited in Related Work, including VoxFormer, FB-Occ, SelfOcc, and FlashOcc. If the claim is restricted to Gaussian-based methods, this should be stated more precisely. More importantly, VG3T, the closest baseline and the main comparison, appears to be the authors' own prior work, and the reported gain over it is 0.78 IoU / 0.29 mIoU with no error bars or multiple-run statistics. Please include the missing baselines or justify their exclusion, and report variance.","section":"Table I / Abstract"},{"comment":"The component activation probability q_{r,k}=1-exp(-beta a_{r,k}) depends only on the amplitude, while the interval probabilities p_{r,i}=1-exp(-Lambda_{r,i}) depend on integrated intensity over a bin. The relationship is not established: a narrow, high-amplitude component can have small integrated intensity but high activation probability, producing Gaussians that are inconsistent with the ray occupancy output. Specify how a_{r,k} relates to Lambda_{r,i}, or define q from the integrated bin probability, so that the adaptive sampling follows the Poisson formulation.","section":"Section III-C, Eq. (7)"}],"minor_comments":[{"comment":"The abbreviations 'Perc.' and 'Dist.' are used without definitions in the caption or text; please define them at first use.","section":"Table II"},{"comment":"The mIoU formula uses C' without explicitly stating that the average is over non-empty classes and that unknown/empty classes are excluded; a clarifying sentence would help reproducibility.","section":"Eq. (16)"},{"comment":"K=3 and beta=1 are fixed hyperparameters but no sensitivity analysis is reported. Given that the method's flexibility depends on these choices, a brief study (e.g., K=1,2,4 or beta in {0.5,2}) would strengthen the robustness argument.","section":"Section IV-C"},{"comment":"The probabilistic Gaussian superposition operator is only cited to GaussianFormer-2; a short equation or precise reference to the formulation would make the rasterization step self-contained.","section":"Section III-D"}],"recommendation":"major_revision","confidential_remarks":"The manuscript builds heavily on VG3T, which appears to be the authors' own prior work, and the reported improvement over it is small and unaccompanied by variance. For a journal submission, I would like to see the controlled ablation isolating the multi-label intensity formulation and an independent strong baseline or error bars. The paper has a good core idea, but the evidence currently underdetermines the causal claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is worth taking seriously. RayOcc replaces the usual softmax-over-depth with a non-normalized Gaussian mixture intensity and maps integrated intensity to bin-wise occupancy via a Poisson link. That is a genuinely different way to model multi-surface rays, and the math in Eqs. (2)-(6) is clean and explicit. The paper also does a nice job of motivating why normalized categorical depth is a poor fit for volumetric occupancy, and the adaptive component-wise sampling is a sensible way to turn mixture amplitudes into variable numbers of Gaussian primitives. I buy the argument that this should help with occlusion and thin structures.\n\nWhat the paper does well experimentally is show that component-wise sampling places Gaussians much closer to occupied voxels than random, categorical, or single-depth initialization (Table II, 61% vs. 28%/16% in Perc.). That is a concrete, reproducible result. The ablation in Table IV also shows the method is not overly sensitive to the activation threshold, which is reassuring.\n\nBut the central claim — that the multi-label Poisson formulation itself drives the SOTA occupancy numbers — is underdetermined. There is no ablation that swaps Eq. (6) for a normalized categorical distribution while keeping everything else fixed. The comparison against VG3T is the closest thing, and the gain there is small (0.78 IoU, 0.29 mIoU), with no error bars. VG3T differs in backbone training, depth regression, and the auxiliary ray loss, so the improvement could come from any of those. The SOTA claim is also narrow: Table I excludes VoxFormer, FB-Occ, SelfOcc, and FlashOcc, so it's really only SOTA among the Gaussian methods they chose to include.\n\nThe ray label generation is another soft spot. Sampling one point at the center of each ~0.55 m bin to label the whole bin is crude and could systematically mislabel thin structures near bin boundaries. Since these labels supervise the mixture intensity and the Gaussian initialization, a labeling bias would propagate directly into the claimed gains.\n\nAll of this is addressable. The idea is sound enough to deserve referee time, but the paper needs a cleaner isolation of the formulation, a broader baseline set, error bars, and fuller implementation details before the claims are accepted.","headline":"RayOcc has a clean non-normalized Gaussian-mixture ray formulation, but the paper doesn't isolate that formulation in experiments; the small gain over VG3T could come from other changes.","tokens_in":12298,"tokens_out":1709,"would_cite":false,"duration_ms":17417,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Camera-only 3D occupancy improves by modeling each ray as a non-normalized Gaussian mixture intensity, letting multiple occluded surfaces coexist per pixel, reaching 34.84 IoU on nuScenes.","keywords":["3D semantic occupancy prediction","camera-only perception","ray occupancy","Gaussian mixture intensity","Poisson event model","multi-label existence","Gaussian splatting","nuScenes"],"falsifier":"Re-derive ray labels by sampling multiple points per depth bin (e.g., 5 or 9 stratified samples) and retrain RayOcc with the same settings; if the IoU and the 61% placement metric shift significantly, the reported improvements depend on the single-sample labeling assumption rather than on multi-hypothesis modeling.","tokens_in":11451,"feed_emoji":"🚗","tokens_out":6936,"duration_ms":59308,"temperature":0.7,"pith_summary":"The paper argues that camera-only 3D semantic occupancy prediction is held back by a formulation mismatch: methods inherited from depth estimation force each ray to choose one dominant depth, while occupancy ground truth is multi-label — a ray can pass through several occupied surfaces. RayOcc replaces the normalized depth distribution with a non-normalized Gaussian mixture intensity along each ray and converts the integrated intensity in each depth bin into an existence probability using a Poisson event model. Because nothing is normalized, multiple depth bins along a ray can all be occupied with high probability without competing. Those mixture components then initialize an adaptive number of 3D Gaussian primitives per ray, which are refined and rasterized into voxel occupancy. On nuScenes the method reports the best IoU (34.84) and mIoU (22.03) among compared Gaussian-based methods, and places 61% of initialized Gaussians inside occupied voxels, up from about 51% for the previous best.","feed_headline":"Multi-hypothesis rays lift camera-only occupancy to 34.84 IoU","feed_subtitle":"Modeling each ray as a non-normalized Gaussian mixture lets one pixel hold several occluded surfaces at once, sharpening 3D occupancy on nuS","key_machinery":"The load-bearing object is the non-normalized Gaussian mixture intensity function λ_r(d) = Σ a_k N(d; μ_k, σ_k²) with no constraint on the amplitudes, together with the Poisson event conversion p_i = 1 − exp(−∫λ). The mixture allows multiple occupancy hypotheses to carry high intensity simultaneously; the exponential map turns integrated intensity into a per-bin existence probability that does not sum to one; and adaptive component-wise sampling (Bernoulli gate with probability 1 − exp(−βa)) converts predicted components into a variable number of 3D Gaussian primitives, which are refined by a sparse convolutional network and rasterized through probabilistic splatting to produce the semantic","core_discovery":"The paper claims that occupancy along a camera ray is a multi-label existence problem, and that the right model is a non-normalized Gaussian mixture intensity rather than a categorical depth distribution. A Mixture Intensity Network predicts K components per ray; the integrated intensity in each depth bin is fed through 1 − exp(−Λ), giving bin-wise existence probabilities that do not compete. Component amplitudes become activation probabilities to spawn 3D Gaussian primitives at multiple depths, refined and rasterized into voxel occupancy. On nuScenes this reaches IoU 34.84 and mIoU 22.03 among compared Gaussian-based methods, with 61% of initialized Gaussians inside occupied voxels.","pith_inferences":["A one-point-per-bin labeling scheme likely mislabels thin structures and bin-boundary surfaces; retraining with multi-point or volume-based ray labels could either strengthen or erode the reported gains, and would clarify whether the mixture model is genuinely learning multi-surface existence.","The paper's advantage over single-depth baselines should be most pronounced in highly occluded urban scenes and should shrink on sparse scenes; a per-scene occlusion-difficulty analysis would provide a direct test of the mechanism.","The connection to NeRF-style transmittance is left largely implicit; using the same intensity for self-supervised depth or appearance prediction could extend RayOcc to settings without dense occupancy labels.","The stochastic Bernoulli gate used in training with deterministic thresholding at inference suggests a Gumbel-softmax reparameterization might deliver lower-variance gradients and further reduce the number of primitives needed."],"forward_implications":["If the formulation is correct, any camera-based occupancy pipeline that currently predicts a softmax depth distribution can be switched to a non-normalized intensity head without architectural changes, gaining multi-surface support for free.","The Poisson conversion introduces no extra trainable parameters, so the method should generalize to other multi-label per-ray problems such as transparent object layout or multi-surface depth estimation.","Because the threshold sweep (τ = 0.7–0.9) leaves occupancy accuracy nearly flat, the model can be tuned at deployment to trade primitive count against latency with little accuracy cost.","The efficiency comparison suggests multi-hypothesis initialization plus sparse refinement can run faster and use less memory than single-hypothesis iterative methods, pointing toward real-time camera-only occupancy."],"fun_headline_variants":["Multi-label ray occupancy via Gaussian mixture intensity","RayOcc: non-normalized mixture models occluded depths per ray","Poisson-form intensity converts rays to multi-surface occupancy","Occlusion-aware rays: 34.84 IoU via Gaussian mixtures"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The ground-truth ray label for each depth bin is obtained by sampling a single 3D point at the bin center and querying a 0.5 m voxel grid; if that point misses an occupied voxel or lands near a boundary, the label is wrong, and that noise supervises the mixture intensity that the entire method rests on.","fun_headline_variants_meta":{"raw":{"variants":["Multi-label ray occupancy via Gaussian mixture intensity","RayOcc: non-normalized mixture models occluded depths per ray","Poisson-form intensity converts rays to multi-surface occupancy","Occlusion-aware rays: 34.84 IoU via Gaussian mixtures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000759,"raw_usage":{"total_tokens":3204,"prompt_tokens":735,"completion_tokens":2469,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":2409}},"tokens_in":479,"tokens_out":2469,"duration_ms":18519,"temperature":1.0,"reasoning_tokens":2409,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T17:21:10.601781+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-derive ray labels by sampling multiple points per depth bin (e.g., 5 or 9 stratified samples) and retrain RayOcc with the same settings; if the IoU and the 61% placement metric shift significantly, the reported improvements depend on the single-sample labeling assumption rather than on multi-hypothesis modeling.","supporting_citations":[],"review_version":1}