{"id":"e6d530a9-e047-42fb-8d12-77dd5604259b","arxiv_id":"2412.11241","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A KDE-based depth outlier rejection step improves RGB segmentation mask precision and downstream panoptic volumetric mapping quality.","lead":"Researchers refine 2D object segmentation masks with depth statistics before building 3D maps, using kernel density estimation to cut depth outliers. The method is tested on a synthetic dataset and a real Baxter robot, but the reported gains come from a comparison that changes both the segmentation model and the refinement step.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table II's system comparison is confounded: PanMap rows use Detectron masks while the proposed rows use a retrained YOLO model, so the 53.7-point gain is not attributable to KDE refinement; only the within-model +10.7 IOU gain is controlled.","rationale":"The reader already reached REJECT and flagged in the rationale that the benchmark changes segmentation models between conditions. I agree that this confound is the most load-bearing problem for the central comparative claim. I do not single out Algorithm 1's tail/nonzero-density assumption as the decisive issue: that limitation is real and should be fixed, but it concerns the method's generality on multimodal depth distributions and is not needed to see that Table II does not support the 'better than PanMap' conclusion. The controlled within-model comparison (rows c vs d) is the only evidence for the KDE contribution, and it is credible as far as it goes. However, because the system-level comparison is confounded and no quantitative 3D metric is reported, the paper's broader claim that the proposed pipeline yields better panoptic volumetric maps than PanMap is not established. A revision that adds the crossed ablation and a volumetric IoU metric could make the claim testable. Since this reinforces rather than changes the reader's verdict, I mark the verdict as unchanged.","tokens_in":9924,"tokens_out":10013,"duration_ms":87657,"concrete_test":"Run a full 2x2 ablation on the flat dataset: (i) PanMap pipeline with raw Detectron masks, (ii) PanMap pipeline with retrained YOLO masks, (iii) proposed pipeline with raw Detectron masks, and (iv) proposed pipeline with retrained YOLO masks; in each condition report both per-frame mask IOU and volumetric IoU against the ground-truth mesh. If condition (ii) reaches roughly the same mask/volumetric IoU as condition (iv), the reported advantage over PanMap comes from the segmentation model rather than from KDE refinement; if (iv) still beats (ii) by a margin comparable to the 10.7-point IOU gain, the refinement claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is that the proposed KDE-based mask refinement, combined with projective SDF mapping, produces better panoptic maps than the prior PanMap system [23]. The evidence for this is Table II. Rows (a) and (b) are produced using the original PanMap segmentation model (Detectron); rows (c) and (d) are produced after retraining the segmentation model with YOLO [37], as stated in Sec. V-B. Consequently, the 53.6577-point difference between row (a) and row (c) is caused primarily by the segmentation model swap, not by the KDE refinement. The only controlled comparison is rows (c) vs (d), which shows a 10.7217 mask-IOU gain from applying Algorithm 1 with a fixed YOLO model. That controlled gain is real but does not establish superiority over PanMap, because no condition runs PanMap with the same YOLO masks. Moreover, the quantitative metric is 2D mask IOU, not any volumetric reconstruction score, so the title claim about volumetric mapping quality rests on qualitative images only. For the central claim to hold, the comparison must isolate the refinement contribution while holding the segmentation model fixed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a non-parametric depth-outlier rejection step for panoptic segmentation masks used in RGB-D volumetric mapping. For each predicted object mask, the depth values under the mask are converted to a 1D kernel density estimate, and depth pixels lying outside the estimated support are removed; the refined masks are then integrated into a projective signed-distance-function (SDF) volume. The method is evaluated on a synthetic 'flat' dataset against the PanMap baseline and demonstrated qualitatively on a Baxter robot with an Intel RealSense camera. The central claim is that this refinement, together with projective SDF mapping, produces more accurate panoptic volumetric maps than the prior PanMap system.","tokens_in":10174,"tokens_out":4887,"duration_ms":46378,"significance":"If the reported gain were properly isolated, the method would be a lightweight and useful addition to RGB-D panoptic mapping: it avoids point-cloud processing, uses a data-driven bandwidth selection, and is accompanied by a public code release. The controlled within-model comparison (rows c vs d of Table II) indicates that the KDE refinement alone improves mask IOU by about 10.7 points on the synthetic sequence, which is a meaningful signal. However, the paper's comparative claim against PanMap is not supported by the experiments as presented, because the segmentation model is changed between conditions, and the only quantitative metric is 2D mask IOU rather than any volumetric reconstruction score.","major_comments":[{"comment":"The central comparison against PanMap is confounded. Rows (a) and (b) use the original PanMap segmentation model (Detectron), while rows (c) and (d) use a retrained YOLO model, as stated in Sec. V-B. The 53.6577-point difference between rows (a) and (c) is therefore attributable primarily to the segmentation-model swap, not to the proposed KDE refinement. The only controlled comparison is rows (c) versus (d), which shows a real 10.7217-point IOU gain from the refinement, but this does not establish superiority over PanMap because no condition runs PanMap with the same YOLO masks. To support the paper's claim, please add a same-segmenter comparison against PanMap (e.g., PanMap with YOLO masks, with and without the proposed refinement) or otherwise isolate the refinement contribution from the segmentation-model change.","section":"Sec. V-B, Table II"},{"comment":"The algorithm assumes that the KDE of each object's depth distribution has one dominant peak and falls below a fixed threshold of 1e-6 on both sides. When the object depth overlaps with background depth, or when the object contains multiple depth modes, the density may never reach that threshold, so the find_id operations either return an empty index or pick an index at the boundary. In that case the algorithm either discards valid object pixels or fails because no cutoff exists. This failure mode is not discussed in the paper, even though it is central to the claimed adaptivity of the method. In addition, the threshold 1e-6 is a manually fixed hyperparameter, which contradicts the abstract's 'without the need for additional parameters' and Sec. IV-C's 'without requiring predefined thresholds'.","section":"Algorithm 1, lines 6-10"},{"comment":"The only quantitative metric reported is 2D mask IOU between the predicted masks and ground-truth annotations. This metric measures the quality of the input masks, not the quality of the resulting volumetric map. The title and abstract claim improvements in volumetric mapping, but the volumetric evidence is qualitative (Figs. 4 and 5) and the real-robot experiments (Sec. VI) are also qualitative. Please either report a volumetric reconstruction metric (e.g., voxel precision/recall, map IOU, or per-object F-score) or restrict the quantitative claim to the mask-refinement stage.","section":"Sec. V, Table II"}],"minor_comments":[{"comment":"The repository link 'https://github.com/mkhangg/refined panoptic mapping' contains a space and is not a valid URL; please provide the correct encoded or actual link.","section":"Abstract and Sec. I"},{"comment":"The symbol M is used for both the input list of masks and the output list; line 2 overwrites the input. Please rename one of them to avoid confusion.","section":"Algorithm 1"},{"comment":"The hole-filling equation does not specify the normalization of the Gaussian weights or the behavior when all neighboring depth values are zero; please clarify the exact update rule.","section":"Eq. (1), Sec. IV-A"},{"comment":"Table I lists 'on-robot real-time performance' for all systems, but the paper reports no timing or computational cost measurements; either provide such measurements or remove/qualify this claim.","section":"Table I, Sec. VI"},{"comment":"The caption says the outliers are 'encoded by the same colors as Fig. 3', but Fig. 3 does not appear to use the same color encoding; please check the cross-reference.","section":"Fig. 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The confounding in Table II is the decisive issue. I would be willing to reconsider a revision that adds a controlled comparison against PanMap with the same segmentation model and a volumetric metric. The algorithm's failure mode when no density cutoff exists should also be addressed explicitly. The current evidence is promising but does not justify the comparative claim as written."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This one is a modest but genuinely new recipe: take the depth values inside a predicted segmentation mask, estimate their 1D density with KDE, and cut the tails. The controlled comparison (your rows c vs d) shows +10.7 mask IOU when the segmenter is held fixed, and that is a real signal. The paper also ships code and describes the pipeline clearly enough that the idea is easy to test once the algorithm is fixed.\n\nBut the paper is not ready as written. Algorithm 1 has an indexing bug: line 6 sets `peak = max(y_kde)` (a density value), then lines 7–8 slice `y_kde[:peak]` and `y_kde[peak:]` as if `peak` were an index. It should be `argmax`. On top of that, line 4 zeroes out depth pixels outside the mask and then flattens the entire depth map, so the KDE sees a huge zero-mass that can dominate the peak. The description in Sec. IV-B says depth values are selected according to the mask, but the pseudocode does not do that. And the 'parametric-free' claim is overstated: the 1e-6 cutoff is a fixed threshold, and if the density never drops below it (background depth overlapping object depth, or multi-modal objects), the algorithm just fails.\n\nThe evaluation is the bigger problem. Table II compares PanMap with Detectron masks against the proposed method with retrained YOLO masks. That 53.7-point gap is mostly the segmenter swap, not the refinement. The only clean comparison is rows (c)–(d), and the paper should present that as the headline result, plus a proper ablation that runs PanMap with the same YOLO masks. The quantitative metric is also 2D mask IOU; the volumetric claim rests on qualitative figures. The real-robot section is illustrative only.\n\nWhat the paper does well: the KDE refinement idea is simple, cheap, and plausibly useful; the within-model gain is meaningful; and the authors correctly position it relative to depth-driven region growing and graph-cut methods. The writing is clear.\n\nBottom line: the idea deserves a serious referee, but the current version needs major revision—fix the pseudocode, correct the zero-filling, run a controlled comparison with a fixed segmenter, and report a volumetric metric. As it stands I wouldn't cite it, and I wouldn't accept it at a journal or conference.","headline":"A simple KDE-based mask refinement idea with a real but confounded evaluation and a broken algorithm listing.","tokens_in":10701,"tokens_out":3266,"would_cite":false,"duration_ms":29860,"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":"This paper proposes a mask-refinement step for panoptic volumetric mapping that cleans up RGB-based segmentation before 3D reconstruction, using kernel density estimation on depth values to cut away over-covered outliers without extra…","keywords":["panoptic mapping","kernel density estimation","depth outlier rejection","segmentation mask refinement","signed distance functions","volumetric mapping","mobile robots","RGB-D perception"],"falsifier":"Run Algorithm 1 on a sequence where a chair is viewed from an angle so its seat and back produce two depth modes and a background wall sits at a similar distance as one mode; if the density between modes never falls below $10^{-6}$, the method will either delete the chair's valid pixels or fail to produce cutoffs, while a manual trim of the same mask would keep the object intact.","tokens_in":9702,"feed_emoji":"🤖","tokens_out":6008,"duration_ms":51924,"temperature":0.7,"pith_summary":"This paper proposes a mask-refinement step for panoptic volumetric mapping that cleans up RGB-based segmentation before 3D reconstruction. The idea is to take each predicted object mask, look at the distribution of depth values it covers, and cut off the low- and high-depth tails of that distribution using kernel density estimation, treating them as segmentation outliers. Because the cutoff is found from the density itself, the method needs no new trained parameters and adapts to out-of-distribution scenes. On a synthetic indoor dataset the refinement improved mask intersection-over-union from 79.8860 to 90.6077 (about ten percent) for a retrained segmenter, and from 16.5150 to 26.2283 for the prior PanMap pipeline. The same pipeline was demonstrated on a real Baxter robot with an Intel RealSense camera, producing cleaner object instances in the resulting SDF map.","feed_headline":"Depth-density fix lifts panoptic map IOU by 10 points","feed_subtitle":"Kernel-density trimming of over-broad object masks sharpens 3D scene maps and runs on a real robot.","key_machinery":"The engine of the method is a one-dimensional kernel density estimate of the depth values covered by each predicted object mask, computed in $O(M \\log M)$ time via FFT-based convolution of grid counts with a Gaussian kernel whose bandwidth is set by the ISJ algorithm. Algorithm 1 takes the density's maximum peak as the object's dominant depth, then scans outward on both sides for the first index where the density drops below $10^{-6}$; those indices become the low and high depth cutoffs. Only depth pixels between the cutoffs survive, and the binary mask is set to zero wherever depth was discarded. This single mechanism does all the work of outlier rejection, and it is what the paper claims makes the refinement adaptive without extra parameters.","core_discovery":"The central claim is that per-instance depth statistics, rather than learned RGB-D fusion, can repair the over-coverage errors that lightweight segmentation networks make. For each binary mask the method computes a kernel density estimate of the depths underneath it, locates the dominant density peak, and sets low and high cutoffs where the density falls below $10^{-6}$; depth pixels outside these cutoffs are removed and the mask is updated accordingly. This is parametric-free in the sense that the bandwidth comes from the ISJ rule and no threshold is tuned per scene. The refined masks then feed a projective signed-distance-function mapper that updates voxels with per-point semantic labels. The paper reports that this refinement raises mask IOU from 79.8860 to 90.6077 on the flat dataset when combined with a retrained YOLO segmenter, and also lifts the prior PanMap system from 16.5150 to 26.2283, indicating the refinement contributes roughly a ten-percent gain on top of whatever segmenter is used.","pith_inferences":["Beyond the paper's claims: the same density-tail criterion could be applied to per-instance color or normal statistics, not just depth, potentially refining masks in settings where depth is noisy or missing.","Beyond the paper's claims: the roughly ten-point IOU gain at both base levels suggests the refinement step is somewhat orthogonal to segmenter quality; a direct test would be to run Alg. 1 with several off-the-shelf panoptic segmenters on the same sequence.","Beyond the paper's claims: the $10^{-6}$ density floor is the one residual magic number; a data-driven or per-object adaptive floor would make the method fully parameter-free and could be tested against the current fixed value."],"forward_implications":["Adding the KDE refinement to an existing RGB segmenter improves final mask IOU by roughly ten percent, independent of whether the base segmenter is weak or strong.","Volumetric maps built with projective SDFs inherit the cleaner masks, so object instances appear with fewer stray point-cloud outliers and sharper boundaries.","Because the refinement step is parametric-free, it can be dropped into a robot's perception stack without retraining or scene-specific threshold tuning.","The same pipeline runs on a real mobile robot (Baxter with an Intel RealSense D435i), so the gains are not limited to synthetic evaluation.","The method is designed to handle out-of-distribution scenes where learned RGB-D fusion degrades, since it relies on the observed depth distribution rather than training-data priors."],"supporting_citations":[{"why":"The SDF-based panoptic mapping system whose masks and maps are the baselines compared in Table II and Fig. 4.","marker":"[23]"},{"why":"Source of the signed distance function volumetric integration used in Eq. 5.","marker":"[32]"},{"why":"The incremental Euclidean SDF mapping framework that the projective SDF update builds on.","marker":"[20]"},{"why":"Earlier volumetric instance-aware semantic mapping that motivates per-instance mask refinement.","marker":"[21]"},{"why":"Foundational real-time dense SDF surface mapping that the reconstruction pipeline inherits.","marker":"[18]"},{"why":"The segmentation model that was retrained to produce the improved base masks in Table II.","marker":"[37]"},{"why":"Defines panoptic segmentation, the task the paper evaluates.","marker":"[27]"}],"fun_headline_variants":["Kernel density fix lifts panoptic map IOU by 10 points","Depth-KDE refinement trims masks, adds 10 IOU points","Panoptic mapping gets 10-point IOU boost via KDE mask repair","Non-parametric depth filter sharpens 3D panoptic maps","Robot 3D mapping: KDE-based mask refinement raises IOU"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The refinement works only when the object's depth distribution has a single dominant peak and drops to near zero on both sides, so if background depth overlaps the object or the object spans two clear depth bands, the algorithm can discard valid object pixels or find no cutoffs at all.","fun_headline_variants_meta":{"raw":{"variants":["Kernel density fix lifts panoptic map IOU by 10 points","Depth-KDE refinement trims masks, adds 10 IOU points","Panoptic mapping gets 10-point IOU boost via KDE mask repair","Non-parametric depth filter sharpens 3D panoptic maps","Robot 3D mapping: KDE-based mask refinement raises IOU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1478,"prompt_tokens":972,"completion_tokens":506,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":410}},"tokens_in":588,"tokens_out":506,"duration_ms":4903,"temperature":1.0,"reasoning_tokens":410,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:08:38.455000+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a sequence where a chair is viewed from an angle so its seat and back produce two depth modes and a background wall sits at a similar distance as one mode; if the density between modes never falls below $10^{-6}$, the method will either delete the chair's valid pixels or fail to produce cutoffs, while a manual trim of the same mask would keep the object intact.","supporting_citations":[{"cited_title":"Panoptic multi-tsdfs: a flexible repre- sentation for online multi-resolution volumetric mapping and long- term dynamic scene consistency,","cited_arxiv_id":null,"evidence_quote":"The SDF-based panoptic mapping system whose masks and maps are the baselines compared in Table II and Fig. 4."},{"cited_title":"A volumetric method for building complex models from range images,","cited_arxiv_id":null,"evidence_quote":"Source of the signed distance function volumetric integration used in Eq. 5."},{"cited_title":"V olumetric instance-aware semantic mapping and 3d object discovery,","cited_arxiv_id":null,"evidence_quote":"Earlier volumetric instance-aware semantic mapping that motivates per-instance mask refinement."},{"cited_title":"Panoptic segmentation,","cited_arxiv_id":null,"evidence_quote":"Defines panoptic segmentation, the task the paper evaluates."}],"review_version":1}