{"id":"74421ab9-7444-493e-a3c7-db3a5ad9f7ca","arxiv_id":"2509.03211","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"ActiveLO-training selects diverse and hard LiDAR sequences so a deep odometry model trained on 52% of the data performs as well as or better than one trained on the full dataset.","lead":"A new training-data selection method lets deep LiDAR odometry models train on roughly half the usual sequences while matching or beating full-dataset accuracy, including in snowy conditions. It first chooses diverse easy-weather sequences, then repeatedly adds the hardest snowy sequences based on the current model's errors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The full-data 100% baseline is described only as 50 epochs with no convergence or variance evidence, so ActiveLO's parity at 52% may reflect a weak reference rather than selection quality.","rationale":"I read the central claim as an empirical statement: a model trained on a 52% ActiveLO-selected subset performs at least as well on Ford, CADC, and WADS test sequences as a model trained on the full 69-sequence pool. For this statement to be meaningful, the full pool model must be a strong, fairly trained reference. The paper provides no evidence that 50 epochs is an adequate or optimal budget for the full model, and the fact that even 23% ActiveLO beats full-data training (Table I) suggests the comparison baseline may be weak. This is a concrete, testable concern rather than a general reproducibility complaint. The reader's weakest-assumption choice, the ITSS feature validity, is real but less decisive because the initial set contains only 6 of 36 selected sequences; if ITSS were a poor heuristic, AIS could still drive most of the benefit, and the 52% claim could survive. By contrast, a weak full-data baseline directly undermines the abstract's headline comparison. I therefore keep the reader's CONDITIONAL verdict unchanged: the result is plausible and the experiments are internally consistent, but the central claim needs a properly converged, repeated full-data baseline before it can be accepted.","tokens_in":17775,"tokens_out":15658,"duration_ms":152455,"concrete_test":"Train HPPLO-Net on the full 69-sequence pool with the same optimizer but for 80 and 120 epochs and with an early-stopping rule on a held-out split; evaluate on the Table I test sets and report the lowest average trel. Also run the 52% ActiveLO and 52% Random protocols with five random seeds each and report mean and standard deviation. If the best full-model trel is below 1.36, or if ActiveLO's error bars overlap the full/random bars, the headline parity claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in the Abstract and Table I is that ActiveLO-training matches full-dataset performance with 52% of the sequence volume. The only specification of the full baseline is '50 epochs' (Eq. 18, Section IV-G); no learning curve, early-stopping rule, hyperparameter search, or repeated trials are reported for it. Since ActiveLO at 23%, 38%, and 52% all beat this baseline by large margins (Table I: trel 1.68, 1.65, and 1.36 vs 1.70), the most parsimonious reading is that the 100% model is undertrained or overfit to the mixed KITTI/CADC pool. If the full model were trained to a properly converged state, the claimed parity might disappear. The absence of error bars, combined with the paper's own admission that Random occasionally wins for about 23% of per-sequence results, means the headline 1.36-vs-1.70 gap is not known to be statistically significant. This concern is more load-bearing than the ITSS feature heuristic: only 6 of the 36 selected sequences come from ITSS, while 30 are chosen by AIS, so even if Eqs. (8)-(9) were arbitrary, the central parity claim could still hold.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ActiveLO-training, a two-stage active selection strategy for building a training set for a deep LiDAR odometry network (HPPLO-Net). The first stage, ITSS, represents each training trajectory as nodes and edges, computes hand-crafted features such as turning-angle variability, edge-length/speed variability, and outlier proportion, and solves a linear program to pick an initial diverse subset (6 of 69 sequences). The second stage, AIS, iteratively adds the 5 sequences with the highest scene-reconstruction loss and prediction-inconsistency loss, computed with the current model, until reaching 52% of the full training pool. Experiments on Ford, CADC, and WADS test sequences compare ActiveLO with random selection and classical LiDAR odometry baselines, reporting that ActiveLO at 52% achieves average translational drift 1.36% versus 1.70% for the full dataset, and consistently beats random selection. The authors claim this demonstrates training efficiency: full-dataset performance with only 52% of the sequence volume.","tokens_in":17954,"tokens_out":4599,"duration_ms":43709,"significance":"If the empirical claims hold, the paper addresses a relevant practical problem: reducing the training cost of deep LiDAR odometry while maintaining or improving generalization across weather conditions. The evaluation is broader than many active-learning papers for odometry, using held-out test sequences from three datasets and comparing against several classical baselines. The paper also provides ablations for both selection components and an explicit (though currently flawed) efficiency model. The central claim is falsifiable and the experimental protocol is, in principle, reproducible. However, the current evidence is not yet convincing because the full-data baseline may not be converged, the random-selection protocol is ambiguous, and no error bars or repeated trials are reported; these issues directly affect the headline 52% parity claim.","major_comments":[{"comment":"The efficiency arithmetic is internally inconsistent. With Num(Sinit)=6, h=5, and iter=7, Eq. (19) evaluates to 6×15 + Σ_{itr=1}^{8} (6+5×itr)×5 = 1230, not 1000 as reported. Eq. (20) evaluates to (69−6) + Σ_{itr=1}^{7} (69−6−5×itr) = 63 + 301 = 364, not 336; the value 336 corresponds to using iter=6 in the sum, contradicting the stated iter=7. The total 1336 in Table VII is therefore also unexplained. Since training efficiency is one of the two claimed contributions, this arithmetic must be corrected and the definition of each term clarified.","section":"§IV-G, Eqs. (18)-(20), Table VII"},{"comment":"The random-selection protocol is ambiguous. The text says that all percentages above 9% 'incorporate the same initial set alongside incremental training data,' but the 9% rows in Tables I-IV report separate ActiveLO and Random results, implying different initial 6-sequence sets. If the 23%-52% ActiveLO results use the ITSS-selected initial set while random results use a different random initial set, the comparison conflates the quality of the initial selection with the quality of the incremental AIS selection. The authors should specify exactly which initial set is used for each row and, ideally, run the incremental process from the same initial set for both strategies to isolate the contribution of AIS.","section":"§IV-B, Tables I-IV"},{"comment":"The full-dataset 100% baseline is described only as '50 epochs' with no convergence criterion, learning curve, or repeated trials. Table I shows that ActiveLO at 23%, 38%, and 52% all outperform the 100% baseline on average trel (1.68, 1.65, and 1.36 versus 1.70), which is more naturally read as evidence that the 100% model is undertrained or poorly regularized than as an indication that 52% of the data is sufficient. The central claim in the Abstract, that ActiveLO 'matches' full-dataset performance, is therefore not established. The authors should report the training curve of the full model, a convergence rule, and ideally multiple seeds, and then revise the claim accordingly.","section":"§IV-B, Table I, §IV-G, Eq. (18)"},{"comment":"No error bars or statistical tests are provided for any of the main comparisons. The paper itself acknowledges that the Random strategy occasionally beats ActiveLO in about 23% of per-sequence results, and Table III shows several sequences (e.g., 54 and 56 at 52%) where random selection has lower trel than ActiveLO. Without repeated runs or a paired test over test sequences, the reported average gaps, such as the 1.36-vs-1.70 headline difference, cannot be distinguished from noise. The authors should add standard deviations or confidence intervals over at least several random-selection seeds and, if feasible, over model-training seeds.","section":"§IV-B, Tables I-IV, §IV-E"},{"comment":"The ITSS objective depends on hand-crafted features with weighting factors λ1 through λ5, but neither the chosen values nor any sensitivity analysis is reported. The only validation of ITSS is the four-sequence ablation in Table V, which compares three fixed random sets against one ITSS set. Given that the 9% ActiveLO row in Table I shows a very large improvement over 9% Random (trel 3.37 versus 7.74), the initial-selection component is load-bearing for the overall framework, and the claim that ITSS captures 'motion diversity' needs stronger support: report the λ values, test nearby values, compare against a standard diversity-based selection baseline, or show that the main result persists with a random initial set when AIS is used.","section":"§III-A.2, Eqs. (8)-(9), Table V"}],"minor_comments":[{"comment":"Steps 3-4 contain typographical errors: 'Nodes and edges enk−1nk used to partition the trajectory nk' should refer to the sequence s, and Eq. references such as '(Eq.(1)˜(2)' are missing closing parentheses.","section":"Algorithm 1"},{"comment":"The objective function sums over 's=0' to '10', but the number of candidate sequences in the pool is not 10; the summation range should be over all sequences in Sge. Please clarify.","section":"§III-A.3, Eq. (10)"},{"comment":"The column labeled 'epochs' actually reports total sequence-epoch iterations, not a number of epochs; the label is misleading and should be changed.","section":"Table VII"},{"comment":"The sentence 'The replicability of our results across multiple experiments further validates the reliability of ActiveLO-training' is not supported by any experimental detail in the paper; either describe the multiple experiments or remove the sentence.","section":"§IV-B"},{"comment":"The distinction between 'Full A-LOAM' and 'A-LOAM' should be explained in the text; currently the reader must infer that 'Full A-LOAM' uses a different training set or preprocessing, and the tables do not state what 'Full' refers to.","section":"Tables II-IV"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core idea is reasonable, but the experimental evidence for the central claim is currently fragile: the 100% baseline appears underconverged, the random-baseline protocol is not fully specified, and the efficiency numbers contain a concrete arithmetic error. These issues are fixable with additional experiments and careful reporting, so I recommend major revision rather than rejection. I also note that the paper's use of HPPLO-Net as the backbone is natural given the authors' prior work, and I do not see a circularity problem because the final evaluation is on held-out sequences."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the headline result, matching full-data training with 52% of sequence volume, is not yet credible because the 100% baseline is under-specified and likely undertrained. That is the load-bearing issue. What is actually new: the two-stage selection pipeline (ITSS plus AIS) for deep LiDAR odometry. The specific combination is, as far as I know, novel. The ablations show both stages contribute, and the experiments cover KITTI, CADC, WADS, and Ford. Credit where due: the incremental selection via scene-reconstruction and prediction-inconsistency losses is a sensible active-learning approach, and the improvement over random selection is consistent across Tables I-IV. The use of HPPLO-Net as the backbone is fine, and self-citing that backbone is not itself a flaw. Soft spots: (1) The full baseline is described only as 50 epochs, with no learning curve, early stopping, hyperparameter search, or repeated trials. ActiveLO at 23%, 38%, and 52% all beat that baseline by large margins, which strongly suggests the 100% model is not converged; parity might vanish against a properly trained reference. (2) There are no error bars anywhere. The paper admits Random sometimes wins on individual sequences, so the headline 1.36-vs-1.70 gap is not known to be statistically significant. (3) Reproducibility is poor: no code or data, and the paper omits the node-selection rule, the number of linear-programming intervals, all weights (lambda1 through lambda5, alpha, beta, epsilon), and the augmentation variances. (4) Table VII's efficiency arithmetic is internally inconsistent: Eq. (19) with iter+1 gives 1230, not 1000, and Eq. (20) with iter=7 gives 364, not 336. The text seems to use iter=7 for training but iter=6 for inference. (5) The 9% random baseline protocol is ambiguous: does it use the same initial six sequences or a fully random draw? None of this kills the paper. The AIS stage selects 30 of the 36 sequences, so even if ITSS's hand-crafted features were arbitrary, the central parity claim could still hold. The core idea is coherent, the empirical trend is promising, and the missing details are fixable. Recommendation: send this to a serious peer review, but the authors must add a properly converged full-data baseline, report error bars, and release code or at least full hyperparameters. I would not cite it in its current form, but I would read the revised version.","headline":"Parity claim at 52% is shaky because the full-data baseline looks undertrained; the two-stage active selection is novel but reproducibility is missing.","tokens_in":789,"tokens_out":2083,"would_cite":false,"duration_ms":34497,"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 two-stage active selection strategy—initial diversity-focused selection plus iterative hard-sample discovery—lets a deep LiDAR odometry model match or beat full-dataset accuracy using only 52% of training…","keywords":["LiDAR odometry","active learning","training set selection","sequence diversity","adverse weather","snow","unsupervised learning"],"falsifier":"Run the same HPPLO-Net training protocol at 52% with three ways of choosing sequences: ActiveLO as described, random subsets averaged over several seeds, and a generic diversity baseline such as farthest-point sampling in feature space. If either baseline matches or beats the reported 1.36% average translation error, the ITSS/AIS scores are not carrying the result. A more direct check is to correlate each sequence's ITSS + AIS score with the reduction in test error obtained by adding that sequence; a near-zero correlation would falsify the selection mechanism.","tokens_in":17438,"feed_emoji":"🚗","tokens_out":11138,"duration_ms":94746,"temperature":0.7,"pith_summary":"Deep LiDAR odometry models estimate vehicle motion from point clouds, and making them work in rain, snow, and dense traffic normally requires large, multi-condition training sets. This paper proposes ActiveLO-training, a two-stage procedure that decides which sequences deserve training time instead of using everything: it scores clear-weather sequences by how much the vehicle's trajectory varies, then iteratively adds the snowy sequences the current model finds hardest, as measured by reconstruction error and prediction inconsistency. The central experimental claim is that a model trained on about 52% of the available sequences matches or slightly beats the same model trained on all of them, with average translational drift of 1.36% versus 1.70% and rotation drift of 0.86°/100m versus 1.00°/100m across the Ford, CADC, and WADS test sets. If this holds, odometry models for new environments can be trained at roughly half the data and compute cost, and rare adverse-weather cases can be deliberately added instead of being left to chance.","feed_headline":"Active sample selection matches full-data LiDAR odometry at 52%","feed_subtitle":"Two-stage selection reaches full-dataset pose accuracy with 52% of sequences, across clear, dynamic, and snowy scenes","key_machinery":"The carrying mechanism is a pair of selection scores. The first is trajectory diversity: per sequence the model computes node turn-angle standard deviation $\\sigma_\\theta$, edge-length standard deviation $\\sigma_l$, edge-speed standard deviation $\\sigma_v$, and outlier proportion, then forms $F_{\\mathrm{Var}}(s)=\\lambda_1\\sigma_\\theta+\\lambda_2\\sigma_l+\\lambda_3\\sigma_v$ and an importance term $F_{\\mathrm{Impor}}(s)$ that rewards total turning and trajectory length. A linear program maximizes $F_{\\mathrm{Var}}+F_{\\mathrm{Impor}}$ subject to sampling at least one sequence from each bin of outlier proportion and average speed. The second score drives incremental selection: for each remaining sequence, the current model computes a point-to-plane scene reconstruction loss $F_{\\mathrm{Recon}}$ plus a prediction inconsistency loss $F_{\\mathrm{Incon}}$ obtained by perturbing the target frame with Gaussian rotations and translations and measuring the variance of the recovered pose; the top $h$ sequences by $F_{\\mathrm{Recon}}+F_{\\mathrm{Incon}}$ enter the training pool each round. These two scores convert the raw point-cloud pool into the 52%-sized training set that the experiments compare against full-data training.","core_discovery":"On the paper's own terms, the discovery is that dataset selection, not network architecture, is the lever for making deep LiDAR odometry robust to diverse weather. ActiveLO-training splits selection into two stages. Initial Training Set Selection (ITSS) converts each trajectory into nodes and edges, computes turn-angle, edge-length, speed, and outlier-proportion statistics, and solves a small linear program whose objective combines trajectory variability and trajectory importance to pick a compact initial set from clear-weather KITTI sequences. Active Incremental Selection (AIS) then iterates: train the current model, run it on the remaining pool, score each sequence by scene reconstruction loss (point-to-plane alignment error) plus prediction inconsistency loss (variance of poses recovered under small random pose augmentations), and add the highest-scoring sequences. In the main tables, the 52%-volume ActiveLO training set yields average translation error 1.36% and rotation error 0.86°/100m, versus 1.70% and 1.00°/100m for the full 100% set, and it beats random selection at every percentage bucket. The authors conclude that a small, deliberately chosen subset can carry the same odometry knowledge as the full pool.","pith_inferences":["An untested stress test: replace ITSS's handcrafted features with a generic diversity baseline such as farthest-point sampling over trajectory feature vectors; if that baseline matches the 52% result, the specific weighting in Eqs. (8)-(9) is not the active ingredient.","The reported experiments use one unsupervised backbone (HPPLO-Net); whether the selection logic transfers to supervised LiDAR odometry networks or to other sensors is left open by the paper.","A deployment-oriented comparison would track wall-clock time, including the per-iteration inference over the remaining pool; sequence-level counts in Table VII describe training load but not the full time budget."],"forward_implications":["A 52%-volume ActiveLO subset produces lower average drift than the full dataset (1.36% vs 1.70% translation error), so the selected set is not merely equal but a better use of the data.","Even at 9% of the data, ActiveLO's ITSS initial set cuts average translation error from 7.74% for random selection to 3.37%, showing a small deliberately diverse starting set already yields a usable base model before snow data is added.","The SRL and PIL ablation (Table VI) shows each loss alone is weaker than the pair, because SRL tends to flag heavy-snow/dynamic-object scenes while PIL flags sparse scenes; keeping both is an explicit design decision supported by the experiments.","Because both stages use trajectory statistics and self-supervised losses rather than labels, the pipeline can rank unlabeled LiDAR logs, which is the form in which odometry training data usually accumulates."],"supporting_citations":[{"why":"Supplies HPPLO-Net, the unsupervised LiDAR odometry backbone used for all ActiveLO training and evaluation runs.","marker":"[32]"},{"why":"A-LOAM is the classical odometry baseline that the 52% ActiveLO model is compared against on the Ford, CADC, and WADS test sets.","marker":"[33]"},{"why":"Provides the point-to-point and point-to-plane ICP baselines (ICP-po2po, ICP-po2pl) used in the snowy-scene comparisons.","marker":"[34]"},{"why":"VGICP is the voxelized GICP baseline benchmarked alongside ActiveLO on the dynamic and snowy sequences.","marker":"[35]"},{"why":"Anchors the point-to-plane alignment error that becomes the Scene Reconstruction Loss used in the AIS stage.","marker":"[28]"}],"fun_headline_variants":["Active selection matches full LiDAR odometry with 52% data","52% of data, full accuracy: active LiDAR odometry training","Selective training beats full set: LiDAR odometry at 52%","Active data pick matches full-dataset LiDAR odometry","Smarter data, same accuracy: 52% LiDAR odometry training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire active-selection argument rests on the assumption that a sequence's trajectory statistics—how much the vehicle turns, changes speed, and produces outliers—measure what makes it valuable for odometry training; if those statistics do not track actual odometry difficulty, ITSS is just a mildly informed heuristic.","fun_headline_variants_meta":{"raw":{"variants":["Active selection matches full LiDAR odometry with 52% data","52% of data, full accuracy: active LiDAR odometry training","Selective training beats full set: LiDAR odometry at 52%","Active data pick matches full-dataset LiDAR odometry","Smarter data, same accuracy: 52% LiDAR odometry training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2935,"prompt_tokens":1019,"completion_tokens":1916,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":1822}},"tokens_in":635,"tokens_out":1916,"duration_ms":12741,"temperature":1.0,"reasoning_tokens":1822,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:33:34.281817+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same HPPLO-Net training protocol at 52% with three ways of choosing sequences: ActiveLO as described, random subsets averaged over several seeds, and a generic diversity baseline such as farthest-point sampling in feature space. If either baseline matches or beats the reported 1.36% average translation error, the ITSS/AIS scores are not carrying the result. A more direct check is to correlate each sequence's ITSS + AIS score with the reduction in test error obtained by adding that sequence; a near-zero correlation would falsify the selection mechanism.","supporting_citations":[{"cited_title":"Hpplo-net: Unsupervised lidar odometry using a hierarchical point-to-plane solver,","cited_arxiv_id":null,"evidence_quote":"Supplies HPPLO-Net, the unsupervised LiDAR odometry backbone used for all ActiveLO training and evaluation runs."},{"cited_title":"A-loam: Advanced implementation of loam,","cited_arxiv_id":null,"evidence_quote":"A-LOAM is the classical odometry baseline that the 52% ActiveLO model is compared against on the Ford, CADC, and WADS test sets."},{"cited_title":"V oxelized gicp for fast and accurate 3d point cloud registration,","cited_arxiv_id":null,"evidence_quote":"VGICP is the voxelized GICP baseline benchmarked alongside ActiveLO on the dynamic and snowy sequences."},{"cited_title":"Linear least-squares optimization for point-to-plane icp surface registration,","cited_arxiv_id":null,"evidence_quote":"Anchors the point-to-plane alignment error that becomes the Scene Reconstruction Loss used in the AIS stage."}],"review_version":2}