{"id":"c32925c3-b63b-4ef1-adaa-c259d6631fb8","arxiv_id":"2411.17432","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A cooperative SLAM and moving-object tracking system selects the most useful neighboring vehicles and the most relevant data to share, cutting communication volume while matching or slightly beating prior methods.","lead":"This paper describes a system for connected self-driving cars that shares only a small amount of selected data between vehicles instead of raw sensor streams, to keep communication costs low while preserving accuracy. It combines place recognition for choosing which vehicle helps with localization and spatial confidence maps for choosing which vehicles help with detecting moving objects.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Communication savings may be an artifact of omitting the two-stage selection handshake (global descriptors, spatial confidence maps) from the communication-volume metric; the paper reports no end-to-end byte accounting.","rationale":"The reader's weakest assumptions (confidence-map reliability and 2D-to-3D zero-padding) identify genuine component-level risks, but even if those were fixed, the headline claim about communication efficiency would still rest on an unverified measurement. The most load-bearing assumption is that the communication-volume metric reflects the actual bytes exchanged by the full system. The paper's own protocol description (Sections III-A, III-B, III-C) shows handshake and broadcast messages that precede the selected content, yet Table IV's CommV ol is inherited from Where2comm without a precise definition for CE CoBEVT and without an end-to-end byte accounting. Without this, the primary quantitative result—substantially lower communication at matched or better detection performance—could be an artifact of an incomplete numerator. This is more central than the confidence-map concern because even a perfectly calibrated confidence map would not repair an inflated efficiency claim. The concern does not overturn the conditional verdict: the architecture is plausible, the comparative AP tables are informative, and the missing byte-level accounting is an obtainable experimental addition. Keeping CONDITIONAL and requiring the byte-level accounting before accepting the efficiency claim as stated is the appropriate disposition. I partially agree with the reader's weakest_assumption: both concerns are legitimate, but the communication-accounting gap is the one that directly supports the paper's title contribution.","tokens_in":17783,"tokens_out":10098,"duration_ms":94155,"concrete_test":"Re-run the CE C-SLAMMOT and CoBEVT pipelines on OPV2V (Default Towns and Culver City) and V2V4Real with packet-level instrumentation that sums every byte transmitted per ego-vehicle update: global descriptors, spatial confidence maps, local feature descriptors, selected sparse feature maps, and all headers/ACKs, across all neighbors. Report total bytes per update for the proposed method and for CoBEVT/Wherr2comm under identical scenarios, and recompute the Table IV trade-off using these totals. If the end-to-end total for CE C-SLAMMOT is not below CoBEVT's total at matched AP@0.5 by at least the claimed margin, or if the request overhead exceeds the selected-feature savings, the communication-efficiency claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is a trade-off between performance and communication cost, but the communication metric may not count all bytes the proposed protocol actually sends. Section III-A describes a two-stage interaction: the ego-vehicle first sends the global descriptor to all neighbors and sends the generated spatial confidence maps to neighbors before receiving selected local features. Table IV's CommV ol follows Where2comm [18], whose metric historically measures transmitted feature messages rather than all handshake or broadcast traffic. The paper never states what CommV ol includes for CE CoBEVT and provides no full-system byte accounting; the only absolute SLAM numbers (~250-286 KB per vehicle) omit the descriptor request and confidence-map broadcast overhead. If the spatial confidence map is an HxW float tensor per neighbor and the global descriptor is broadcast to all candidates, the real per-update communication can be substantially larger than the reported selected-feature volume. The claimed '32.6% less' than CoBEVT at equal AP@0.5 is computed from the CommV ol values in Table IV, so it inherits any such omission. This is an internal measurement gap, not a disagreement with consensus, and it directly gates the headline efficiency result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents CE C-SLAMMOT, a cooperative SLAMMOT system that reduces communication by (i) using a sequence global descriptor (SeqVPR) and local features (SuperPoint/SuperGlue) to select a single best collaboration vehicle for inter-vehicle relative pose estimation, and (ii) using spatial confidence maps (from Where2comm) to select which neighboring vehicles and which spatial regions should be shared for cooperative object detection, fusing the selected features with a CoBEVT-style axial attention. The system is evaluated on OPV2V and V2V4Real, comparing ego-pose accuracy, detection AP, MOTA, and communication volume against SLAM-only, cooperative SLAM, C-SLAMMOT, and cooperative perception baselines. The reported results show that CE C-SLAMMOT achieves ego-pose errors within a few centimeters of the raw-data C-SLAMMOT baseline, and CE CoBEVT reaches or exceeds CoBEVT accuracy at lower reported communication volume.","tokens_in":17967,"tokens_out":8439,"duration_ms":77064,"significance":"If the reported results hold, the paper provides evidence that content-aware, confidence-guided selection can replace raw observation sharing in tightly coupled cooperative SLAMMOT without significant accuracy loss, and that the perception module advances the state of the art on two public benchmarks. The work integrates two previously separate research threads, and the experiments use public datasets and published baselines, which strengthens external validity. The proposed CE CoBEVT module and the unified factor-graph backend are concrete contributions. However, the headline efficiency result is currently gated by measurement and reporting gaps, so the significance depends on whether those gaps can be closed.","major_comments":[{"comment":"The communication-volume metric is not defined end-to-end, and the headline reductions do not follow from the table. Section III-A describes a two-stage handshake: the ego vehicle broadcasts a global descriptor to all neighbors and sends spatial confidence maps before receiving selected local features or feature maps. The absolute SLAM numbers in Section IV-B (250-286 KB) count only local features from the selected vehicle, and Table IV's CommVol is adopted from Where2comm without stating whether it includes the descriptor broadcast and confidence-map messages. The claimed savings (32.6% on OPV2V Culver City, 6.5% on Default Towns, 16.1% on V2V4Real) also cannot be read off Table IV, because no CE CoBEVT row at the matched AP exists for those percentages. Please provide a full byte-level accounting for the complete protocol, including control messages, and specify exactly how the percentage savings are computed, for example by interpolation on the AP-CommVol curve.","section":"III-A, Table IV"},{"comment":"No error bars, confidence intervals, or significance tests are reported. The central claim that selective communication maintains accuracy comparable to full communication rests on point estimates whose differences are often very small, e.g., MEAN 0.226 versus 0.219 in the OPV2V T-junction in Table III. Without variance across runs, scenes, or resamples, the reader cannot judge whether the observed differences are noise. Please add repeated-run or per-scene variability and, where possible, paired significance tests for the main comparisons, especially for Table III, Table IV, and Table VII.","section":"Tables III-VII"},{"comment":"The relative pose estimation converts a 2D BEV transformation into 3D by zero-padding, which sets z, roll, and pitch to zero. This assumption is load-bearing for the inter-vehicle factor in the graph optimization, but the paper provides no justification or sensitivity analysis. On non-flat or banked roads the zero-padding could bias the ego-pose estimates. Please either validate that the test scenes are sufficiently planar, quantify the error introduced by the 2D assumption, or extend the method to estimate the full 3D transformation.","section":"III-B3, Eqs. (1)-(2)"},{"comment":"The selection mechanism for cooperative perception assumes that the ego spatial confidence map, with request map N = 1 - C, reliably indicates where perception is deficient. If the ego detector is confidently wrong, for example due to a missed occluded object or a hallucinated object, the request map will direct attention to irrelevant areas and the performance-communication trade-off will degrade. No calibration or failure analysis of the confidence map is provided. Please add a discussion of this failure mode and an experiment that quantifies degradation when the confidence map is imperfect.","section":"III-C"}],"minor_comments":[{"comment":"The table header contains the typo \"sptical\", which should be \"spatial\". Also, the text mentions a communication rate of 0.808 without defining how this rate is calculated.","section":"Table VII"},{"comment":"Equation (8) has duplicated superscripts in the first sum, and the accompanying text says \"standard derivation\" where \"standard deviation\" is intended; please fix the equation formatting and the wording.","section":"Eq. (8)"},{"comment":"The table caption should clarify that the six rows for each method group correspond to different datasets (Default Towns, Culver City, V2V4Real), as the current layout is easy to misread.","section":"Table IV"},{"comment":"The text says Figure 3 shows results with different numbers of collaboration vehicles, but the caption only lists two scenes and does not explain how the additional vehicles are selected from the ranking; please clarify the selection procedure for the multi-vehicle cases.","section":"IV-B, Fig. 3"},{"comment":"The paper states that pre-trained SuperPoint weights are used on BEV intensity images without discussing whether this introduces a domain gap; a sentence on why the transfer is reasonable, or a comparison with a BEV-trained feature extractor, would help.","section":"III-C"},{"comment":"The title and abstract claim that the method determines the number of collaboration vehicles, but the two modules select numbers independently and the conclusion states that finding a unique number when both aspects are considered is future work; this scope limitation should be acknowledged earlier in the paper.","section":"Abstract, Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper combines two known ideas (Where2comm-style confidence maps and CoBEVT fusion; descriptor-based place recognition) into a unified C-SLAMMOT framework, which is a reasonable incremental contribution for the journal. The main gate for publication is the communication-accounting issue: the headline savings must be verifiable from a clearly defined, end-to-end byte metric. The second gate is the absence of any uncertainty quantification in the experimental tables; for a paper whose core claim is that performance is preserved while communication is reduced, point estimates without variance are insufficient. I would not reject the paper, but the revision needs to address both points substantively."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper does something genuinely useful and the experiments mostly back it up, but the headline claim about communication savings is not actually measured. The core idea, picking one vehicle for cooperative SLAM via sequence descriptors and a separate variable set for perception via spatial confidence maps, is a sensible extension of the authors' earlier C-SLAMMOT and of Where2comm. Tables III to VII broadly show accuracy within a few centimeters or a few AP points of the raw-data baseline. Credit where due: the integration is coherent, the evaluation covers two public datasets plus noisy-pose and tracking variants, and the comparison set is fair.\n\nThe soft spots are real, and one is load-bearing. The paper's communication metric does not count its own handshake traffic. The two-stage protocol first broadcasts the global descriptor to all neighbors and the spatial confidence maps to all neighbors, then receives selected local features. The reported SLAM message size (about 250 to 286 KB) is just the local features, and Table IV's CommV ol follows Where2comm, which typically counts transmitted feature messages only. No end-to-end byte accounting is given. So the reductions like \"32.6% less than CoBEVT\" are at best partial. That number is arithmetically consistent with the table (23 times 0.674 is about 15.5, which is where CE CoBEVT matches CoBEVT's AP@0.5), so the reader's suspicion of a numeric mismatch is not supported; the problem is the omitted overhead, not the arithmetic. A serious revision should report total bytes, including the request messages, and show the trade-off both with and without them.\n\nTwo smaller assumptions deserve flagging. The 2D-to-3D relative pose via zero-padding assumes planar ground motion; fine for these datasets, but it should be stated. And the request map N = 1 - C inherits Where2comm's trust in the ego detector's confidence; confidently-wrong detections would request the wrong content. No calibration analysis is provided.\n\nThe title overpromises: the method \"determines\" one SLAM partner and a variable perception set, but joint selection is explicitly left to future work. Also, no code, data, or error bars are provided, so the small performance differences are not significance-tested. None of these kills the work; they just limit how strongly it can be cited.\n\nVerdict: this deserves peer review, not a desk reject. A referee should focus the authors on the byte accounting, the planar-motion assumption, and error bars. If the handshake overhead turns out to be large, the central claim will need to be rescaled, but the system is still a reasonable contribution to C-SLAMMOT.","headline":"Useful integration of selective communication into C-SLAMMOT, but the headline bandwidth savings are unmeasured because the protocol's handshake traffic isn't counted.","tokens_in":18565,"tokens_out":4430,"would_cite":true,"duration_ms":56205,"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":"The paper claims that a connected vehicle fleet can match raw-data cooperative SLAMMOT accuracy while sending only compact descriptors and spatially sparse feature maps, because each vehicle can determine which partners and which spatial…","keywords":["cooperative SLAM","multi-object tracking","communication efficiency","spatial confidence map","place recognition","relative pose estimation","LiDAR perception","vehicle-to-vehicle collaboration"],"falsifier":"Use a scene where the ego detector is confidently wrong, for instance a fully occluded vehicle it places with high confidence or a hallucinated object in an empty region. If, under that confidence map, the request mechanism fails to request the occluded region and detection or tracking accuracy collapses to single-vehicle level while communication stays low, then the confidence-map assumption is the point of failure.","tokens_in":17457,"feed_emoji":"🚗","tokens_out":11195,"duration_ms":101072,"temperature":0.7,"pith_summary":"Cooperative SLAMMOT, which simultaneously localizes a vehicle, builds a map, and tracks moving objects with help from other vehicles, normally requires every car to broadcast its raw LiDAR scan to every partner, which does not scale. The paper claims that the same accuracy is reachable by first sending two tiny request messages: a sequence-level global descriptor that identifies the one partner with the most similar map, and a spatial confidence map that marks where the ego vehicle's own detector is unsure. Only the chosen partners then send back local feature descriptors and spatially sparse, confidence-weighted feature maps, which are fused in a joint factor-graph backend. In experiments on simulated and real cooperative-driving datasets, ego-pose error stays within a few centimeters of the raw-sharing baseline while communication volume drops by up to a third, and the cooperative detector beats the previous best detector at equal or lower bandwidth. The value of the claim, if true, is that cooperative perception and localization become practical under realistic bandwidth limits as fleet size grows.","feed_headline":"Selective sharing matches raw cooperative driving at lower bandwidth","feed_subtitle":"Sending only descriptors and sparse maps, it picks the partners that matter while keeping pose and detection accuracy.","key_machinery":"The load-bearing mechanism is a two-stage request-and-response selection scheme. In the cooperative SLAM stage, a sequence-level global descriptor, a compact vector encoding spatio-temporal features of recent bird's-eye-view intensity frames, is broadcast as the request; the responding vehicle is the one whose local map matches best, and it sends only local feature descriptors for relative pose estimation. In the cooperative perception stage, the request is the spatial confidence map $C$, inverted to $N = 1 - C$, which scores where the ego detector is unsure; partners are ranked by how little their detected occupied regions overlap the ego's, and each sends only the confidence-weighted sparse feature map in the requested areas. The selected relative-pose factor and object-perception factors are then optimized together with odometry factors in a joint factor-graph backend, with asynchronous incorporation of non-keyframe object detections. The key identity is the complementarity rule $V_{e\\rightarrow n} = f_{\\mathrm{decide1}}(\\mathrm{sum}(D_e \\cap D_n)/\\mathrm{sum}(D_e \\cup D_n))$, which converts two confidence maps into a decision about whether a neighbor's information is complementary enough to request.","core_discovery":"The paper's central claim is that content-aware selection of collaboration vehicles, rather than raw observation sharing, is sufficient to preserve the benefits of tightly coupled cooperative SLAM and moving-object tracking. For localization, the ego vehicle broadcasts only a sequence global descriptor built from its recent bird's-eye-view intensity images; the neighbor with the most similar local map is chosen, and only that vehicle's local feature descriptors and current pose are transmitted, letting the ego solve an inter-vehicle relative pose by feature matching and RANSAC. For perception, the ego's spatial confidence map is inverted into a request map $N = 1 - C$, and partners whose detected occupied regions have the least overlap with the ego's are selected; each selected partner sends only the spatially sparse feature-map cells that the ego requested, and these are fused through an axial attention module. A joint factor graph then integrates the resulting relative-pose factor and object-perception factors with odometry and motion factors, and the reported experiments show this system matching the raw-data C-SLAMMOT baseline in ego-pose MEAN/RMSE, exceeding the previous best cooperative detector on the OPV2V and V2V4Real benchmarks, and cutting communication volume while keeping detection accuracy.","pith_inferences":["The same request-map logic could be applied to heterogeneous sensor suites: a vehicle with cameras could request specific bird's-eye-view regions from a LiDAR-equipped partner, since the request is expressed in spatial confidence rather than raw sensor format.","A natural follow-up is to make the confidence map end-to-end trainable with a communication penalty, so the system learns where to be uncertain rather than treating the detector's confidence as fixed.","The two selection mechanisms suggest a unified utility: score each potential partner by a weighted combination of place-recognition similarity and spatial complementarity, which could determine a single number of collaboration vehicles serving both localization and perception simultaneously.","The observed improvement over full-map fusion hints at a broader principle: gating out irrelevant information can act as a regularizer for cooperative perception, not just a bandwidth saver."],"forward_implications":["Cooperative localization and perception can be decoupled from raw bandwidth: the same pose accuracy as raw sharing is reachable by exchanging compact descriptors and sparse features, so the number of collaborating vehicles is no longer limited by link capacity.","Dynamic vehicle selection introduces a tunable performance-communication knob, allowing operators to set a target confidence or accuracy and let the system choose how many vehicles to contact rather than always contacting all.","Because perception requests are spatial, a partner only needs to send the regions the ego lacks, which naturally suppresses irrelevant or distracting information and can improve detection relative to full-map fusion.","The two selection criteria, place similarity for SLAM and spatial complementarity for perception, can be tuned independently, so a fleet can optimize localization partners and perception partners separately."],"supporting_citations":[{"why":"Baseline C-SLAMMOT that shares raw point clouds among all vehicles; the paper compares against it directly and extends its factor-graph backend.","marker":"[16]"},{"why":"Introduces the spatial confidence map and the request-map idea ($N=1-C$) used to choose complementary vehicles and spatial areas.","marker":"[18]"},{"why":"Supplies the fused axial attention module and the previous best cooperative perception baseline that the new detector module is compared against and improves.","marker":"[10]"},{"why":"Provides the sequence-level global descriptor method used for place recognition, which selects the collaboration vehicle in the cooperative SLAM module.","marker":"[31]"},{"why":"The learned local feature extractor applied to bird's-eye-view intensity images for inter-vehicle relative pose estimation.","marker":"[34]"},{"why":"The learned feature matcher that, with RANSAC, turns local feature pairs into the inter-vehicle relative pose.","marker":"[41]"},{"why":"OPV2V simulation benchmark used for ego-pose, detection, and tracking comparisons.","marker":"[42]"},{"why":"V2V4Real real-world benchmark used to test ego-pose and detection under realistic conditions.","marker":"[43]"},{"why":"Lightweight LiDAR SLAM backend that generates the single-vehicle odometry factors entering the graph optimization.","marker":"[38]"},{"why":"Asynchronous object-state estimation scheme that lets non-keyframe detections contribute to the joint factor graph.","marker":"[45]"}],"fun_headline_variants":["Choosing key vehicles matches raw C-SLAMMOT at lower bandwidth","Smart partner selection cuts communication without losing accuracy","C-SLAMMOT: sharing only critical data keeps performance","Cooperative SLAMMOT selects partners to save bandwidth","Descriptor-based sharing matches raw data with less traffic"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the ego vehicle's spatial confidence map reliably shows where its own perception is deficient, so every low-confidence region is actually a region where help is needed.","fun_headline_variants_meta":{"raw":{"variants":["Choosing key vehicles matches raw C-SLAMMOT at lower bandwidth","Smart partner selection cuts communication without losing accuracy","C-SLAMMOT: sharing only critical data keeps performance","Cooperative SLAMMOT selects partners to save bandwidth","Descriptor-based sharing matches raw data with less traffic"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000642,"raw_usage":{"total_tokens":3023,"prompt_tokens":1082,"completion_tokens":1941,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":698,"completion_tokens_details":{"reasoning_tokens":1860}},"tokens_in":698,"tokens_out":1941,"duration_ms":13550,"temperature":1.0,"reasoning_tokens":1860,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:07:16.404061+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use a scene where the ego detector is confidently wrong, for instance a fully occluded vehicle it places with high confidence or a hallucinated object in an empty region. If, under that confidence map, the request mechanism fails to request the occluded region and detection or tracking accuracy collapses to single-vehicle level while communication stays low, then the confidence-map assumption is the point of failure.","supporting_citations":[{"cited_title":"Where2comm: Communication-efficient collaborative perception via spatial confidence maps,","cited_arxiv_id":null,"evidence_quote":"Introduces the spatial confidence map and the request-map idea ($N=1-C$) used to choose complementary vehicles and spatial areas."},{"cited_title":"CoBEVT: Coop- erative bird’s eye view semantic segmentation with sparse transformers,","cited_arxiv_id":null,"evidence_quote":"Supplies the fused axial attention module and the previous best cooperative perception baseline that the new detector module is compared against and improves."},{"cited_title":"Learning sequence descriptor based on spatio-temporal attention for visual place recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the sequence-level global descriptor method used for place recognition, which selects the collaboration vehicle in the cooperative SLAM module."},{"cited_title":"Superpoint: Self- supervised interest point detection and description,","cited_arxiv_id":null,"evidence_quote":"The learned local feature extractor applied to bird's-eye-view intensity images for inter-vehicle relative pose estimation."},{"cited_title":"Superglue: Learning feature matching with graph neural networks,","cited_arxiv_id":null,"evidence_quote":"The learned feature matcher that, with RANSAC, turns local feature pairs into the inter-vehicle relative pose."},{"cited_title":"OPV2V: An open benchmark dataset and fusion pipeline for perception with vehicle-to- vehicle communication,","cited_arxiv_id":null,"evidence_quote":"OPV2V simulation benchmark used for ego-pose, detection, and tracking comparisons."},{"cited_title":"V2V4Real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception,","cited_arxiv_id":null,"evidence_quote":"V2V4Real real-world benchmark used to test ego-pose and detection under realistic conditions."}],"review_version":1}