{"id":"5f1076d3-434f-4c60-a5c0-18fd75c85d1d","arxiv_id":"2501.18110","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A lifelong 3D mapping framework that removes dynamic points, aligns multi-session maps, detects changes, and stores only deltas so any past session map can be reconstructed without raw maps.","lead":"This paper presents a modular system for keeping 3D maps of buildings or streets up to date across repeated LiDAR scans. It removes moving objects, aligns maps from different sessions, detects changes, and uses a version control scheme that stores only differences, so old maps can be reconstructed and any two sessions can be compared without keeping the raw maps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reconstructing previous session maps via Eq. (1) is never quantitatively validated: the convex-hull crop cannot isolate non-convex session extents, and imperfect change-detection deltas are compounded, so the central version-control claim is unsupported.","rationale":"We read the paper as a modular lifelong mapping system whose distinctive contribution is map version control that reconstructs prior session maps without storing them. The reader's weakest assumption matches the most load-bearing concern: Eq. (1) is the only mechanism for reconstruction, and it is not validated against original session maps. We agree with the CONDITIONAL verdict. The paper does provide credible quantitative evaluations for dynamic object removal (Table II) and multi-session alignment (Table III), and the memory-efficiency numbers (Table V) are plausible if reconstruction works. However, the reconstruction claim itself has no direct evidence: no table or figure reports error between reconstructed maps and originals. The convex-hull issue is concrete and structural — base maps intentionally contain older data, and a convex hull is an over-approximation for non-convex session extents. This is not a matter of consensus; it is an internal correctness risk of the central claim. We do not call for rejection because the system may perform adequately in near-convex scenes, and the missing evaluation could be supplied; but the paper as submitted should not be accepted without that quantitative reconstruction check. Hence UNCHANGED from the reader's CONDITIONAL.","tokens_in":13116,"tokens_out":4255,"duration_ms":41485,"concrete_test":"Run the full pipeline on the LT-ParkingLot and NCLT multi-session sequences. For each session k, reconstruct M'_s(k) using Eq. (1) and the stored deltas and boundaries. Compare M'_s(k) to the original clean session map M_s(k) using the symmetric Chamfer distance with tau=0.5 as in Sec. IV-B, plus point-level precision/recall. Then repeat the reconstruction using an alpha-shape or the exact occupied-region boundary instead of the convex HullFilter_k. If Chamfer distance is non-negligible or precision/recall drops materially (e.g., >5%) with the convex hull, the central version-control claim is not supported as stated; if the alpha-shape version is markedly better, the convex-hull crop is the identified cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that map version control reconstructs any previous clean session map from the current base map, stored deltas, and boundary points, without storing raw maps. This relies on Eq. (1) (Sec. III-E), assuming that (a) the stored base_ND and session_PD point sets exactly capture all changes between sessions, and (b) applying HullFilter_k with the session's convex-hull boundary crops the accumulated base map to precisely the session's extent. Both assumptions are insecure. Change detection (Sec. III-D) operates at BEV pixel granularity (0.05-0.15 m), and Table IV reports PD/ND precision as low as 0.769 and recall as low as 0.690; the deltas are not exact point sets, and errors accumulate over the summation in Eq. (1). More structurally, the base map intentionally preserves occluded regions from earlier sessions (Sec. III-A, Fig. 4(b)), so base map(k) contains points never observed in session k. Cropping by the convex hull of session k's boundary cannot remove those spurious interior points when the session's true extent is non-convex (e.g., L-shaped corridors, multi-level structures). The paper never compares reconstructed M'_s(k) with the original session maps, leaving the central version-control functionality unverified and, for non-convex environments, likely to return maps that include points from other sessions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a modular lifelong 3D mapping framework for hand-held and robot-mounted LiDAR mapping. It consists of four components: a dynamic point removal pipeline based on OctoMap, plane regression and k-NN voting; a two-stage multi-session map alignment using PCA-SHOT descriptors, RANSAC and NDT; a BEV-image-based map change detection module that outputs positive and negative differences; and a map version control system that maintains a single base map and stores deltas plus boundary points. The paper claims that this design permits reconstruction of any previous clean session map and querying of inter-session changes without storing the original session maps. Experiments cover SemanticKITTI, NCLT, MulRan, LT-ParkingLot and two hand-held XGrid datasets, with quantitative results for dynamic removal, alignment, change detection and memory efficiency.","tokens_in":13383,"tokens_out":5640,"duration_ms":53764,"significance":"The potential practical value is real: a robust lifelong mapping pipeline that separates dynamic-object removal, alignment, change detection and version control could simplify long-term map maintenance, and the memory-efficiency numbers in Table V are encouraging. The paper is also stronger than many module papers in evaluating each component on multiple real-world datasets and comparing against ERASOR, Removert, Ground-Octomap, ICP, NDT and LT-Mapper. The dynamic-object-removal results on SemanticKITTI reach competitive F1 scores. However, the central novel claim—lossless or near-lossless reconstruction of arbitrary previous session maps through Eq. (1)—is not validated quantitatively at all, and the alignment evaluation uses the test data for parameter selection. These gaps currently prevent the paper from fully supporting its headline functionality.","major_comments":[{"comment":"The central version-control claim is not quantitatively supported. The reconstruction formula M'_s(k)=HullFilter_k{M_b(t)+ sum_i (base_ND(i-1) - session_PD(i))} is never compared with the original session map. This is load-bearing because Table IV shows the delta point sets are not exact: PD precision is as low as 0.769 and ND recall as low as 0.690, so errors accumulate over the summation. In addition, the base map deliberately retains occluded points from earlier sessions (Sec. III-A, Fig. 4(b)), and a convex-hull boundary cannot remove such interior points when a session's true extent is non-convex (e.g., L-shaped corridors or multi-storey structures). At minimum, report quantitative reconstruction error per session (e.g., Chamfer distance and pointwise precision/recall against the original clean session maps) and include at least one non-convex or multi-level environment.","section":"III-E, Eq. (1)"},{"comment":"The multi-session alignment comparison is weakened by parameter selection on the test maps. Section III-C states that the optimal alignment is chosen by selecting the registration with the lowest Chamfer distance, and Table III reports only the resulting average Chamfer distance. This means the reported numbers can reflect fitting to the test data, not a generalizable parameter setting. Please use a held-out validation split for grid-search selection, or report results over repeated runs with error bars and cross-validation, and state explicitly how parameters were chosen for each dataset.","section":"IV-B, Table III"},{"comment":"The quantitative change-detection evaluation uses manually introduced relocations as ground truth, and a detected point is counted as true if a ground-truth change lies in 'its small radial neighborhood.' This tolerance and the manually constructed changes need sensitivity analysis; without reporting the radius and its effect, the precision/recall numbers in Table IV are difficult to interpret. Moreover, because Table IV also shows ND recall of 0.798 (XGrid-Outdoor) and PD recall of 0.690 (XGrid-Parking), the deltas stored by the version control are not exact, reinforcing the need for the reconstruction validation requested above.","section":"IV-C, Table IV"},{"comment":"The sensor-setup-agnostic dynamic removal claim is only quantitatively validated on SemanticKITTI, a robot-mounted automotive setup; the hand-held XGrid results are qualitative. Section III-B itself notes that the optional height filter may not work for drone or multi-storey hand-held mapping and that 1-3% of static points are typically misclassified. To support the hand-held claim, add a quantitative evaluation on hand-held data with annotated or simulated dynamic objects, or explicitly restrict the claim to ground-like single-level hand-held scenarios.","section":"IV-A, Fig. 6"}],"minor_comments":[{"comment":"The abstract says 'without manual parameter fine-tuning' and Sec. III calls the alignment 'automatic parameter-free', but Sec. III-C and Table I describe a grid search over six hyperparameters; please replace 'parameter-free' with 'automatic' or clarify that parameters are selected automatically.","section":"Abstract and Sec. III-C"},{"comment":"Equation (1) and the accompanying example do not specify the index handling for k=t (base_ND(t-1) is not defined) and do not state whether the HullFilter is applied before or after adding deltas; define the notation and boundary cases.","section":"III-E, Eq. (1)"},{"comment":"Typos and wording: 'VERION' in Table V caption, 'very unique' repeated in abstract and introduction, 'Lets look' in Sec. III-E, and 'Negative difference is the objects' in Fig. 3 caption.","section":"Multiple locations"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting and the modular system is well motivated. The main issue for the editor is that the headline 'reconstruct any previous clean session map' is presented in the abstract and introduction but never measured. I would like to see a dedicated reconstruction experiment before acceptance; if the authors cannot provide one, the version-control contribution should be reframed as approximate delta compression with explicit error bounds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the genuinely new idea is delta-based map version control, and it's not in the cited prior work. The integrated pipeline is a competent modular system. But the central claim—reconstructing any previous session map from base map plus deltas—is never quantitatively tested, and the reconstruction equation has a structural weakness that makes the claim doubtful for non-convex environments.\n\nWhat's good: The modules are sensible and the evaluation is broader than most: dynamic removal on SemanticKITTI with a strong F1, alignment results on several datasets with larger Chamfer improvements, change detection precision/recall on four datasets, and memory efficiency up to 94% on NCLT. The paper also correctly positions against LT-Mapper and identifies why single-point-cloud version control doesn't transfer. The writing is clear.\n\nWhere it's soft. The reconstruction claim is load-bearing and unsupported. Eq. (1) assumes the stored positive and negative differences exactly capture all change, and that the convex hull boundary crops the base map to the session's extent. The paper's own change detection numbers show precision and recall well below 1 (e.g., PD precision 0.769 and PD recall 0.690 on some datasets), so the deltas are not exact. Because the base map deliberately preserves occluded regions from prior sessions, a convex hull cannot remove spurious interior points when the session's true extent is non-convex. The paper never compares a reconstructed M_s(k) to the original session map. That is the key missing experiment.\n\nSecondary concerns: dynamic removal's sensor-agnostic claim rests on one quantitative dataset (SemanticKITTI); hand-held results are qualitative. Alignment has no error bars and selects parameters by minimizing Chamfer distance on the test maps, which is a mild selection-bias concern. No code or data release is provided, and several thresholds are unspecified. These are fixable.\n\nOverall, the framework is plausible and the version-control idea is worth pursuing, but the headline functionality is unverified. If I were the editor, I would send it to review rather than desk reject, because the idea is new and the system is substantial; the main ask would be a direct reconstruction evaluation plus code and data. The stress-test note's central concern is correct and should be the focus of that review.","headline":"The delta-based map version control is genuinely new and the integrated system is substantial, but the central reconstruction claim is never validated and is structurally doubtful for non-convex environments.","tokens_in":13934,"tokens_out":2140,"would_cite":false,"duration_ms":22133,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A lifelong 3D mapping framework claims that a single current base map plus stored positive and negative differences can reconstruct any earlier clean LiDAR session map and report changes between any two sessions, without keeping the…","keywords":["lifelong mapping","LiDAR","dynamic object removal","map alignment","map change detection","map version control","point cloud","multi-session mapping"],"falsifier":"Take a multi-session dataset of the kind used in the paper, reconstruct each historical session map using Eq. (1), and compute the mean nearest-neighbor distance (Chamfer distance) between each reconstruction and the original clean session map before the originals are discarded. If the mean distance is comparable to or larger than the map resolution (0.2–0.5 m), or if the convex-hull crop removes or adds visible geometry at the map edges, then the exact-reconstruction claim would be falsified.","tokens_in":12891,"feed_emoji":"🗺️","tokens_out":8027,"duration_ms":77713,"temperature":0.7,"pith_summary":"Long-lived robots and digital twins need maps that stay current as the world changes, but re-storing each new LiDAR session is wasteful. This paper proposes a modular pipeline that takes any sequence of session maps, removes moving objects, aligns them into one reference frame, detects what appeared and disappeared, and then keeps only a single current base map together with the detected differences and per-session boundaries. The central claim is that this compact representation is enough to reconstruct any earlier clean session map on demand and to answer queries about changes between any two sessions, without archiving the original heavy maps. The reported storage saving grows with the number of sessions, reaching about 94% on a 27-session, 1.5-year dataset. If the reconstruction is truly lossless, lifelong mapping becomes a version-control problem instead of a data-retention problem.","feed_headline":"Keep one 3D map, restore any past session","feed_subtitle":"Storing only the current map plus differences cuts storage up to 94% and answers what changed over time.","key_machinery":"The load-bearing object is the reconstruction identity in Eq. (1): $M'_s(k) = \\mathrm{HullFilter}_k\\{M_b(t) + \\sum_{i=t}^{k} (\\mathrm{baseND}(i-1) - \\mathrm{sessionPD}(i))\\}$, where $M_b(t)$ is the current base map, $\\mathrm{baseND}$ and $\\mathrm{sessionPD}$ are the stored negative and positive difference point sets, and $\\mathrm{HullFilter}_k$ crops the accumulated cloud to the convex hull boundary stored for session $k$. That equation is fed by the map change detection module, which uses radius-based nearest-neighbour searches to separate each map into coexisting, overlapping, and non-overlapping parts and bird's-eye-view height descriptors to locate changed pixels, and by a two-stage alignment (feature-descriptor matching followed by a normal-distributions-transform fine registration) that puts sessions in a common frame. All the later claims — storage savings, historical reconstruction, inter-session change queries — hang on this equation being reversible under real alignment errors.","core_discovery":"The paper's central discovery is a way to turn a growing pile of large 3D LiDAR maps into a version-controlled single map. Each new session is cleaned of dynamic points, aligned to the current base map, and compared with it; the differences are classified as negative (present in the base map but absent in the new session) and positive (present in the new session but absent in the base map). The base map is then updated by removing the negative differences and adding the positive ones. The system keeps only the current base map, the saved positive and negative difference sets, and the convex-hull boundary of each session map. Its claim is that Eq. (1) can walk this store backwards: add the saved negative differences back and subtract the positive differences in reverse order, then crop with the requested session's boundary, to reproduce any earlier clean session map exactly enough for practical use — so users can retrieve any historical map or diff any two sessions without the raw session data.","pith_inferences":["The paper leaves implicit that the same delta store could support incremental semantic annotation: because every positive and negative difference is localized, a label added to one session's object could be propagated through time by replaying the deltas instead of re-segmenting every map.","A testable extension is to replace the convex-hull boundary with a tighter crop, such as an alpha shape or a per-session occupancy boundary; non-convex session footprints, like an L-shaped building or a corridor network, would expose whether the hull filter is the accuracy-limiting component of Eq. (1).","The reported 94% saving on the year-long dataset compares stored deltas against downsampled maps; the paper does not report how the same representation behaves when input maps are kept at full resolution, so the scaling of reconstruction error with point density remains an open check.","If reconstruction fidelity holds across dozens of sessions, the framework turns 'what changed over the last year?' into a single replay operation, suggesting a natural product-level query: diff any two historical states without storing either state at full resolution."],"forward_implications":["As more sessions are added, stored data grows only by the size of detected differences, so memory use flattens: the paper reports efficiency rising from about 40% at 3 sessions to 94% at 27 sessions.","Any earlier clean session map can be reconstructed through Eq. (1) without the original raw map, so long-running deployments can archive a compact delta store instead of full-resolution session dumps.","Users can obtain the changes between any two sessions by reconstructing both and running the change detection module, enabling temporal queries such as which objects appeared or disappeared.","Because the dynamic removal step does not assume horizontal LiDAR or planar motion, the same pipeline is claimed to work for hand-held scanners and robot-mounted SLAM systems."],"supporting_citations":[{"why":"OctoMap supplies the occupancy framework that separates initial static, dynamic, and unknown points in the dynamic removal module.","marker":"[10]"},{"why":"ERASOR provides the dynamic-removal baseline and the preservation/rejection/F1 evaluation protocol used on SemanticKITTI.","marker":"[9]"},{"why":"LT-Mapper is the modular lifelong mapping system the framework is compared against and whose alignment and change-detection limitations motivate the new design.","marker":"[1]"},{"why":"The existing point-cloud version control system is the prior art the paper argues cannot handle multi-session maps because point ordering and alignment errors break its assumptions.","marker":"[19]"},{"why":"NDT registration is the fine-alignment stage in the two-stage multi-session map alignment pipeline.","marker":"[25]"},{"why":"SemanticKITTI provides point-wise ground-truth static/dynamic labels for the quantitative dynamic object removal evaluation.","marker":"[28]"}],"fun_headline_variants":["One LiDAR map, version-controlled for life","Reconstruct any past LiDAR session from one map","Cut storage 94% with 3D map version control","One base map, any past session replayable","Version control for lifelong 3D LiDAR maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reconstruction equation assumes that adding and subtracting the saved difference point sets to the current base map, then cropping with a session's convex-hull boundary, exactly restores every earlier map even when session alignment has small errors and the stored difference points are imperfect.","fun_headline_variants_meta":{"raw":{"variants":["One LiDAR map, version-controlled for life","Reconstruct any past LiDAR session from one map","Cut storage 94% with 3D map version control","One base map, any past session replayable","Version control for lifelong 3D LiDAR maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2857,"prompt_tokens":974,"completion_tokens":1883,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":1821}},"tokens_in":590,"tokens_out":1883,"duration_ms":12694,"temperature":1.0,"reasoning_tokens":1821,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T00:37:13.478928+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a multi-session dataset of the kind used in the paper, reconstruct each historical session map using Eq. (1), and compute the mean nearest-neighbor distance (Chamfer distance) between each reconstruction and the original clean session map before the originals are discarded. If the mean distance is comparable to or larger than the map resolution (0.2–0.5 m), or if the convex-hull crop removes or adds visible geometry at the map edges, then the exact-reconstruction claim would be falsified.","supporting_citations":[{"cited_title":"OctoMap: An efficient probabilistic 3D mapping framework based on octrees","cited_arxiv_id":null,"evidence_quote":"OctoMap supplies the occupancy framework that separates initial static, dynamic, and unknown points in the dynamic removal module."},{"cited_title":"ERASOR: Ego- centric Ratio of Pseudo Occupancy-based Dynamic Object Removal for Static 3D Point Cloud Map Building","cited_arxiv_id":null,"evidence_quote":"ERASOR provides the dynamic-removal baseline and the preservation/rejection/F1 evaluation protocol used on SemanticKITTI."},{"cited_title":"LT-mapper: A Modular Framework for LiDAR-based Lifelong Mapping","cited_arxiv_id":null,"evidence_quote":"LT-Mapper is the modular lifelong mapping system the framework is compared against and whose alignment and change-detection limitations motivate the new design."},{"cited_title":"Ogayar-Anguita, Alfonso L ´opez-Ruiz, Rafael J","cited_arxiv_id":null,"evidence_quote":"The existing point-cloud version control system is the prior art the paper argues cannot handle multi-session maps because point ordering and alignment errors break its assumptions."},{"cited_title":"Fast and accurate scan registration through minimization of the distance between compact 3d ndt representations","cited_arxiv_id":null,"evidence_quote":"NDT registration is the fine-alignment stage in the two-stage multi-session map alignment pipeline."},{"cited_title":"Behley, M","cited_arxiv_id":null,"evidence_quote":"SemanticKITTI provides point-wise ground-truth static/dynamic labels for the quantitative dynamic object removal evaluation."}],"review_version":1}