{"id":"3c4dd4a0-f1d7-4d53-be92-38e66ef987ab","arxiv_id":"2608.08696","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"OccAnyScene trains one Gaussian-based occupancy model on indoor and outdoor datasets at once and loses less than half a point of mIoU versus training each separately.","lead":"This paper introduces a new task, cross-scene 3D occupancy prediction, and a Gaussian-based model, OccAnyScene, that jointly handles indoor and outdoor scenes with different cameras, ranges, and voxel grids. It matters because a single vision model that works at both room scale and street scale could simplify robots and autonomous vehicles that move between the two.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-run comparisons cannot establish 'negligible' cross-scene degradation; the 0.41/0.19 mIoU drops may be within seed noise.","rationale":"I read the strongest claim as an empirical stability claim: one model jointly trained on Occ-ScanNet and SurroundOcc-nuScenes performs almost as well as two separately trained models. This rests on four numbers: 59.51 versus 59.92 mIoU (drop 0.41) and 22.87 versus 23.06 mIoU (drop 0.19), both with the DAv3 encoder. The paper provides no measure of run-to-run variability. Because the claimed effect is a small difference, the absence of variance is a direct threat to the conclusion. This is more load-bearing than the pixel-frustum scale formula: although the formula is heuristic, its removal in Table 5 degrades both datasets, so it is at least empirically motivated, whereas the stability comparison itself has no distributional support. The inter-camera-gap module is a legitimate secondary concern, explicitly admitted in the Limitations and omitted from the main method section and Fig. 2, but the paper states that it has little effect on aggregate metrics, which is checkable by ablating it. Given the small observed drops, a multi-seed evaluation would settle whether the stability claim is real or within noise. I therefore keep the reader's CONDITIONAL verdict unchanged, with the additional condition that the authors provide repeated-seed statistics and ideally an ablation of the gap-query module.","tokens_in":14021,"tokens_out":10006,"duration_ms":116815,"concrete_test":"Run OccAnyScene-DAv3 scene-specific and cross-scene training on both datasets with at least three random seeds, fixing all other hyperparameters, and report mean and standard deviation for mIoU and IoU in each configuration. Also report the per-seed difference (cross-scene minus scene-specific) with a confidence interval. If the 95% interval for either dataset contains a drop larger than 1 mIoU, or if the observed drops of 0.41 and 0.19 are within one standard deviation of the difference distribution, the 'negligible degradation' claim is not established and should be reported as within noise. As a secondary check, ablate the supplementary spatial-query Gaussians in the cross-scene nuScenes model to verify the paper's assertion that they have little effect on aggregate metrics.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that joint indoor-outdoor training degrades mIoU by only 0.41 on Occ-ScanNet and 0.19 on SurroundOcc-nuScenes relative to scene-specific training. These differences are small, and the paper reports a single training run per configuration, with no standard deviation, no seed protocol, and no statement about whether the scene-specific and cross-scene runs shared initialization or hyperparameters. On a metric like mIoU, run-to-run variation is often of this magnitude, so the reported 'stability' is not yet distinguished from noise. The reader's candidate weakest assumption, the pixel-frustum scale formula b_{p,k} = eta * 0.5 * (d/fx + d/fy) in Section 3.3, is a legitimate secondary concern, but Table 5 already shows that removing frustum-relative scaling degrades both datasets, so the formula is at least empirically motivated. What is missing is evidence that the headline comparison itself is reproducible. A separate concern explicitly flagged in the Limitations is the unablated spatial-query module for inter-camera gaps, which is absent from Section 3 and Fig. 2; if removing it changes nuScenes mIoU materially, the attribution of outdoor performance to the pixel-frustum representation would need re-scoping. The decisive issue, however, is the lack of repeated-seed evidence for the quantitative stability claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new task setting, Cross-Scene 3D Semantic Occupancy Prediction, in which a single model must handle indoor and outdoor scenes with different cameras, spatial ranges, voxel resolutions, and semantic taxonomies. The authors propose OccAnyScene, which builds on a pretrained depth foundation model and consists of two main modules: Pixel-Aligned Frustum Feature Aggregation (PFFA) constructs camera-aware frustum queries, and Frustum-Parameterized Gaussian Construction (FPGC) decodes each query into Gaussians whose positions and scales are constrained by the pixel frustum geometry. The model is evaluated on Occ-ScanNet (indoor) and SurroundOcc-nuScenes (outdoor). The paper reports state-of-the-art scene-specific results and, as its central contribution, shows that a jointly trained cross-scene model nearly matches separately trained scene-specific models, with mIoU drops of 0.41 on Occ-ScanNet and 0.19 on nuScenes using the DAv3 encoder, whereas an adapted SplatSSC baseline drops 5.03 and 1.19 under the same joint training.","tokens_in":14368,"tokens_out":5051,"duration_ms":51488,"significance":"If the central cross-scene stability claim holds, this is a valuable contribution: it defines a meaningful new task and demonstrates a practical way to consolidate indoor and outdoor occupancy prediction into one model without large accuracy loss. The methodological idea of using the pixel frustum as a geometric reference for Gaussian positions and scales is well motivated and is supported by component and fine-grained ablations (Tables 4 and 5), an efficiency analysis, and qualitative visualizations. The paper is also commendably explicit about its limitations, including the dataset scope and the inter-camera coverage mechanism. The main weakness is that the headline quantitative claim rests on single training runs, with no measure of run-to-run variability; for a claim of 'negligible degradation', that evidence is not yet conclusive.","major_comments":[{"comment":"The central claim that joint cross-scene training causes negligible degradation is based on mIoU differences of 0.41 on Occ-ScanNet and 0.19 on SurroundOcc-nuScenes between the scene-specific and cross-scene models. The paper reports only one training run per configuration, with no standard deviation, no number of seeds, and no statement about whether the scene-specific and cross-scene runs shared initialization or hyperparameters. On a metric like mIoU, run-to-run variation is commonly of this magnitude, so the reported stability is not yet distinguished from noise. Please provide repeated-seed results (at least three runs, with means and standard deviations) for the scene-specific versus cross-scene comparison, or otherwise quantify the expected variability of the differences.","section":"Section 4.3, Tables 2 and 3"},{"comment":"The model includes a supplementary spatial-query module that covers inter-camera gaps in SurroundOcc-nuScenes, but this module is described only in the Limitations section and is absent from the method description and Fig. 2. No ablation quantifies its effect on nuScenes metrics. Since the paper attributes the stability of the cross-scene model to the pixel-frustum representation, and the queries produce the occupancy predictions in camera-unobserved regions, their contribution to the reported nuScenes numbers should be measured. The statement that these queries have little effect on aggregate metrics is an assertion, not evidence; please provide an ablation with and without the module, and if the module is retained, document it in the method section.","section":"Section 5 (Limitations) vs. Section 3 and Fig. 2"}],"minor_comments":[{"comment":"The abstract highlights the scene-specific SOTA numbers (59.92% and 23.06% mIoU) rather than the cross-scene stability numbers, which are the paper's novel contribution. It would be clearer to state the cross-scene results and their near-equivalence to scene-specific models in the abstract.","section":"Abstract and Section 4.3"},{"comment":"The entry '59.1064.40' appears to be a typographical error and likely should be '59.10 64.40' with a separator; the current formatting makes the table difficult to read.","section":"Table 2, row OccAnyScene-DAv3 Cross-scene"},{"comment":"The text refers to 'depth increments' while the table row is labeled 'w/o depth residual'; please use consistent terminology for this component.","section":"Section 4.4, Table 5"},{"comment":"The scale reference b_{p,k} = eta * 0.5 * (d/fx + d/fy) is the key mechanism for cross-scene scale adaptation, but the value of the hyperparameter eta is not reported and no sensitivity analysis is provided. Please state the chosen value and, ideally, show how results vary with eta.","section":"Section 3.3, pixel-frustum-relative scale parameterization"},{"comment":"The implementation details state that the effective number of iterations per dataset is kept identical to the single-dataset setting, but no details are given on learning rate schedules, batch sizes, or how the alternating schedule is implemented; please clarify these for reproducibility.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-engineered and the proposed method is plausible, but the decisive claim of 'negligible' cross-scene degradation needs multi-seed evidence. If repeated runs are not feasible, the wording should be softened accordingly. The supplementary spatial-query module should be moved into the method description and ablated, as its omission is a reproducibility gap. The work fits the journal's scope and is worth a revision round."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is not just another occupancy model: it proposes a concrete new task protocol (single model across indoor and outdoor datasets with different camera geometries, voxel specs, and taxonomies) and a clean geometric trick to make it work — expressing Gaussian positions and scales relative to the pixel frustum cross-section at predicted depth. Second, the paper's central quantitative claim, that joint training costs less than half a point of mIoU on each dataset, is plausible but not yet demonstrated because every number in the paper is a single run with no variance.\n\nWhat is actually new: the pixel-frustum-relative scale formula b = eta*0.5*(d/fx + d/fy) is a sensible, non-obvious way to keep Gaussian sizes comparable across very different cameras and scene ranges. The ablations back it up: removing frustum-relative scale degrades both datasets, and the canonical-camera depth mapping is credited to Metric3Dv2. The paper also does the right thing by wrapping their own SplatSSC baseline with dataset-specific adapters so the cross-scene comparison is apples to apples. The limitations paragraph is unusually candid: it admits the inter-camera gap regions in nuScenes are handled by a separately placed set of spatial queries, which are not described in the method section or Fig. 2 and are never ablated. That is a real exposition gap — if those queries contribute materially to the outdoor numbers, the attribution of stability to the pixel-frustum representation needs to be re-scoped.\n\nThe soft spots, in order. First and most important, the two headline drops (0.41 and 0.19 mIoU) sit right in the usual range of seed noise for a model like this. With no error bars, no seed protocol, and no statement about whether the scene-specific and joint runs share initialization, the claim that cross-scene training is 'negligible' is an estimate, not a result. Second, the abstract leads with scene-specific SOTA numbers rather than the cross-scene numbers that are the paper's actual contribution; that's a presentation choice that undercuts the framing but is easy to fix. Third, the inter-camera query module is essentially invisible outside the limitations. Fourth, the hyperparameter eta is introduced but never given a value or sensitivity check; minor.\n\nIf I were the editor, I would send this to review. The task setting and parameterization deserve referee time, and the core evidence is consistent even if not yet statistically airtight. I would ask the authors to run multiple seeds (or at least 3) for the scene-specific vs cross-scene comparison, and to fold the spatial-query module into the method and ablations. A reader in 3D occupancy or embodied perception should read this; the task protocol alone is worth knowing about.","headline":"A solid, honest occupancy paper with a new task protocol and a good geometric trick, but the headline stability claim needs repeated seeds before it is trusted.","tokens_in":14869,"tokens_out":2243,"would_cite":true,"duration_ms":24481,"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":"One Gaussian occupancy model can be trained jointly on indoor and outdoor scenes without losing accuracy, because each primitive is sized from the camera-dependent pixel frustum at its predicted depth.","keywords":["3D semantic occupancy prediction","cross-scene learning","indoor and outdoor scenes","Gaussian splatting","pixel frustum","depth foundation model","scene-adaptive parameterization","voxel occupancy"],"falsifier":"Retrain the same joint model on a third occupancy dataset with a camera outside the tested range, such as a wide-angle indoor sensor; if the frustum-scale reference drives the stability, the cross-scene mIoU drop should stay near zero, while a miscalibrated reference should reproduce the several-point degradation of the SplatSSC baseline. A cheaper test is to replace $b_{p,k}$ with a single learned constant per dataset and check whether the joint-training gap widens, which would implicate the frustum geometry rather than the depth features.","tokens_in":13835,"feed_emoji":"🧊","tokens_out":11231,"duration_ms":108032,"temperature":0.7,"pith_summary":"The paper introduces Cross-Scene 3D Semantic Occupancy Prediction, a setting in which a single model must predict occupancy for both indoor and outdoor scenes under their native cameras, spatial ranges, voxel sizes, and semantic label sets. The authors argue that the real obstacle is geometric: Gaussian primitives that represent occupied space must grow with depth and change with focal length, so a shared model cannot rely on absolute metric scales that look nothing alike in a 4.8 m room and a 100 m street. OccAnyScene answers this with a pixel-frustum-centered Gaussian representation whose offsets and scales are expressed relative to the metric cross-section of each pixel's frustum at the predicted depth. The reported results show that one jointly trained model stays within 0.41 mIoU and 0.19 mIoU of separately trained scene-specific models on Occ-ScanNet and SurroundOcc-nuScenes, while a prior Gaussian baseline loses 5.03 and 1.19 points under the same joint training. If correct, this means a general-purpose occupancy predictor can serve environments as different as rooms and roads without per-scene retraining or model switching.","feed_headline":"One model now maps both rooms and streets to 3D occupancy","feed_subtitle":"Joint indoor-outdoor training costs just 0.41 mIoU indoors and 0.19 outdoors with the shared Gaussian model.","key_machinery":"The central object is the pixel-frustum-centered Gaussian: a 3D Gaussian primitive whose position and scale are constrained by the frustum of the feature pixel that generates it, where a pixel frustum is the pyramid-shaped region of space subtended by that pixel, widening with depth and depending on camera focal length. It is built by two modules: Pixel-Aligned Frustum Feature Aggregation, which combines per-pixel geometry features, the camera ray direction, and surrounding context through deformable cross-attention to produce a frustum query, and Frustum-Parameterized Gaussian Construction, which decodes each query into K Gaussians. The scale of the k-th Gaussian is $s_{p,k} = b_{p,k} \\hat{s}_{p,k}$, where $b_{p,k} = \\eta \\cdot \\frac12\\,(d_{p,k}/f_x + d_{p,k}/f_y)$ is the metric cross-section width of the pixel frustum at the Gaussian's depth and $\\hat{s}_{p,k}$ is a learned dimensionless relative scale. This lets the network predict only depth, bounded offsets, and dimensionless scales, so the metric outcome recomputes itself for any camera intrinsics and scene range.","core_discovery":"OccAnyScene is a single feed-forward model that predicts 3D semantic occupancy for both indoor Occ-ScanNet and outdoor SurroundOcc-nuScenes protocols, with all weights shared except for dataset-specific taxonomy matrices. It uses a pretrained depth foundation model for visual tokens and dense geometry features, then per feature pixel builds a camera-aware frustum query and decodes it into K Gaussians. The Gaussians are anchored at a predicted metric surface depth, displaced into occluded regions by learned depth increments, shifted laterally by bounded subpixel offsets, and scaled by a dimensionless factor times the frustum's metric width at that depth. The paper reports state-of-the-art scene-specific scores of 59.92% mIoU on Occ-ScanNet and 23.06% mIoU on SurroundOcc-nuScenes, and shows that joint cross-scene training with DepthAnythingV3 costs only 0.41 and 0.19 mIoU respectively, whereas a SplatSSC-based cross-scene baseline drops 5.03 and 1.19 points.","pith_inferences":["The same frustum-relative argument predicts that the model should transfer to other projection models, such as fisheye or narrow-field cameras, once the frustum cross-section formula is replaced by the appropriate projection geometry; the paper does not test this.","A direct way to separate geometric adaptation from feature transfer is to ablate the frustum reference with a single learned scale per dataset; if the cross-scene gap stays small, the depth foundation model's features, not the frustum geometry, are what make the model stable.","The paper evaluates only two datasets, so the strongest version of the claim, that the parameterization generalizes to arbitrary unseen occupancy protocols, remains open; adding a third indoor or outdoor dataset with a very different camera would be the natural next test."],"forward_implications":["A deployed system can serve both road driving and indoor parking with one occupancy model, since only the output taxonomy matrix changes between domains.","Joint training across the two protocols costs only 0.41 and 0.19 mIoU with the DAv3 encoder, indicating that the shared parameterization, not per-scene tuning, is what carries cross-scene stability.","The end-to-end objective optimizes depth and occupancy together in a single loss, removing the two-stage training used by prior Gaussian occupancy methods.","Because the Gaussian scale reference is computed from depth and focal length, the same weights can be evaluated under camera intrinsics different from training without rescaling voxels or re-tuned priors."],"supporting_citations":[{"why":"Supplies the pretrained depth foundation model (ViT encoder plus DPT head) that provides visual tokens and dense geometry features, and is the default encoder for the method.","marker":"[30]"},{"why":"Provides the stronger DepthAnythingV3 encoder used to show that the cross-scene stability holds with a larger model.","marker":"[14]"},{"why":"Supplies the canonical-camera depth prediction scheme that reconciles metric depth across datasets with different focal lengths.","marker":"[8]"},{"why":"Supplies the Decoupled Gaussian Aggregator and occupancy losses used to splat Gaussians onto the target voxel grid, and is the scene-specific baseline the paper adapts for cross-scene comparison.","marker":"[17]"},{"why":"Defines the indoor Occ-ScanNet benchmark whose spatial range, voxel size, and taxonomy set the indoor protocol for joint training.","marker":"[31]"},{"why":"Defines the outdoor SurroundOcc-nuScenes benchmark whose surround-view cameras, range, and taxonomy set the outdoor protocol.","marker":"[26]"},{"why":"Introduces the 3D Gaussian primitive representation that OccAnyScene generalizes from image pixels to scene-adaptive frustum-centered Gaussians.","marker":"[12]"}],"fun_headline_variants":["One model predicts 3D occupancy indoors and out","Single Gaussian framework unifies indoor-outdoor 3D occupancy","OccAnyScene: joint indoor-outdoor 3D occupancy with tiny accuracy cost","Shared weights map rooms and streets to 3D occupancy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that half the sum of depth over the two focal lengths, the metric width of a pixel's frustum at the Gaussian's depth, is the correct unit of Gaussian size for every camera and scene range, and that its base magnitude, set by an undisclosed constant eta, is calibrated well enough to generalize; if this geometric reference is wrong, the shared parameterization fails exactly in the way the paper claims to prevent.","fun_headline_variants_meta":{"raw":{"variants":["One model predicts 3D occupancy indoors and out","Single Gaussian framework unifies indoor-outdoor 3D occupancy","OccAnyScene: joint indoor-outdoor 3D occupancy with tiny accuracy cost","Shared weights map rooms and streets to 3D occupancy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000685,"raw_usage":{"total_tokens":3124,"prompt_tokens":982,"completion_tokens":2142,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":2070}},"tokens_in":598,"tokens_out":2142,"duration_ms":14842,"temperature":1.0,"reasoning_tokens":2070,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:26:18.103654+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same joint model on a third occupancy dataset with a camera outside the tested range, such as a wide-angle indoor sensor; if the frustum-scale reference drives the stability, the cross-scene mIoU drop should stay near zero, while a miscalibrated reference should reproduce the several-point degradation of the SplatSSC baseline. A cheaper test is to replace $b_{p,k}$ with a single learned constant per dataset and check whether the joint-training gap widens, which would implicate the frustum geometry rather than the depth features.","supporting_citations":[{"cited_title":"Splatssc: Decoupled depth-guided gaussian splatting for semantic scene completion","cited_arxiv_id":null,"evidence_quote":"Supplies the Decoupled Gaussian Aggregator and occupancy losses used to splat Gaussians onto the target voxel grid, and is the scene-specific baseline the paper adapts for cross-scene comparison."}],"review_version":1}