{"id":"3a7b5c5c-25e8-4b96-8571-331e85eef22f","arxiv_id":"2412.10033","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TimeAlign uses Swin-LSTM prediction and camera-guided combination of past and observed LiDAR BEV features to partially recover 3D detection accuracy under LiDAR time lag.","lead":"This paper introduces TimeAlign, a module that predicts and fuses LiDAR features from past frames to keep 3D object detection accurate when LiDAR data arrives late relative to cameras. It reports improvements over the GraphBEV baseline on nuScenes under simulated one-frame LiDAR lag, with a small accuracy cost on normal data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TimeAlign's reported gain over GraphBEV under one-frame LiDAR lag is confounded by extra historical LiDAR inputs: the GraphBEV baseline sees one sweep, while TimeAlign sees three or more.","rationale":"The reader's weakest assumption concerns whether the random frame-drop corruption used in training faithfully models real fixed one-frame LiDAR lag. That is a legitimate concern about generalization. However, the more immediate load-bearing issue is that the experimental comparison does not control for input history: TimeAlign receives additional LiDAR sweeps that the GraphBEV baseline does not, so the measured gain cannot be attributed to the proposed modules without a controlled ablation. This is not a question of statistical noise or corruption distribution; it is a confound in the causal claim. The paper's own statement that GraphBEV uses sweeps=1 while TimeAlign uses three historical frames makes the confound explicit. A simple control variant of GraphBEV with the same multi-sweep input would settle the issue. Because the existing CONDITIONAL verdict already requires stronger evidence, the verdict remains UNCHANGED rather than moving to ACCEPT or REJECT.","tokens_in":8755,"tokens_out":4711,"duration_ms":52794,"concrete_test":"Train and evaluate a GraphBEV control that consumes the same T-3..T-1 (or T-3..T) LiDAR sweeps under the one-frame lag protocol, for example by feeding the four sweeps through the existing BEV encoder and combining their BEV features by averaging or summation, with no TimeAlign prediction or combination modules. If this control reaches approximately 0.656 AP-Car on the same test set, the claimed TimeAlign benefit disappears; if it stays near 0.595 AP-Car, the modules are doing real work. Additionally, report results on the full nuScenes validation split with at least three random seeds to rule out mini-test noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison in Table 2 contrasts GraphBEV under LiDAR lag (AP-Car 0.595) with TimeAlign under the same lag (AP-Car 0.656). However, the two models do not receive the same sensor input. GraphBEV is evaluated with settings K=8 and sweeps=1 (Section 4), so under a one-frame LiDAR lag it effectively sees only the delayed sweep at T-1. TimeAlign, by design, consumes a series of three historical LiDAR point clouds plus the current observed frame (Sections 3.1 and 3.2), so under the same lag it has access to T-3, T-2, and T-1 (with T replaced by a sampled previous frame during training). The improvement attributed to the TimeAlign prediction/combination modules is therefore confounded with simply having more LiDAR sweeps as input. Temporal aggregation alone is known to improve detection (Section 2.2 cites BEVDet4D and BEVFusion4D), so a baseline that aggregates the same historical sweeps without any learned prediction could account for part or all of the 0.061 AP-Car gap. The paper provides no ablation that isolates the Swin-LSTM prediction and dual-transformer combination from the effect of adding historical frames.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TimeAlign, a modification of the GraphBEV multi-modal 3D object detector to handle temporal misalignment, specifically LiDAR frame lag relative to the camera stream. The method consumes three historical LiDAR sweeps (T-3 to T-1) plus the currently observed sweep, uses a Swin-LSTM predictor to forecast the current LiDAR BEV feature from the historical frames, and combines the predicted and observed features under the guidance of camera BEV features via a dual-transformer combination module. During training, the current LiDAR input is randomly replaced by a previous frame using a random variable α to simulate LiDAR lag. Experiments on the nuScenes dataset report AP for four object classes with and without one-frame LiDAR lag, comparing against GraphBEV. Under lag, TimeAlign improves car AP from 0.595 to 0.656, but on synchronized data car AP degrades from 0.823 to 0.789. The paper also claims to prove that temporal misalignment is distinct from spatial misalignment, based on a single experiment with GraphBEV.","tokens_in":8981,"tokens_out":5380,"duration_ms":56888,"significance":"If the claimed gains are robust, the work addresses a relevant and under-studied failure mode of multi-modal perception in autonomous driving: asynchronous LiDAR and camera streams. The proposed architecture is a plausible extension of GraphBEV, and the idea of using recurrent prediction and camera-guided combination to recover temporally misaligned LiDAR features is interesting. The training-time frame-drop corruption is a practical contribution that may help models become robust to LiDAR lag. However, the current evidence is not convincing: the main comparison in Table 2 is confounded by unequal sensor input (extra historical sweeps for TimeAlign), the test set is small and unspecified, no error bars or ablations are provided, and the normal-case performance drops noticeably. The central claim is thus defensible but not yet demonstrated.","major_comments":[{"comment":"The main comparison between TimeAlign and GraphBEV under one-frame LiDAR lag is confounded by the number of LiDAR sweeps available to each model. GraphBEV is evaluated with sweeps=1 (Section 4), so under a one-frame lag it sees only the delayed sweep at T-1. TimeAlign, by design, receives three historical sweeps plus the current observed frame (Sections 3.1 and 3.2). Since temporal aggregation alone is known to improve detection (as the paper itself notes in Section 2.2, citing BEVDet4D and BEVFusion4D), the 0.061 AP-Car improvement in Table 2 cannot be attributed solely to the proposed prediction and combination modules. A baseline GraphBEV variant that consumes the same three historical sweeps without the Swin-LSTM prediction or dual-transformer combination is necessary to isolate the method's contribution.","section":"Section 4, Table 2; Sections 3.1, 3.2"},{"comment":"The evaluation is performed on an unspecified 'mini test set' with no sample size, no error bars, and no repeated runs. The nuScenes mini split, if used, typically comprises only a few hundred frames, so the reported differences (e.g., car AP 0.595 vs. 0.656) may be within noise. The authors should report the number of test samples, confidence intervals or multiple-seed variance, and ideally results on the full nuScenes validation set to establish statistical reliability.","section":"Section 4, Table 2"},{"comment":"The statement that 'we proved that temporal misalignment is not consistent with space one' is an overstatement. Table 1 only shows GraphBEV's degradation under temporal lag; it does not include a controlled comparison with spatial misalignment on the same data and settings. The claimed robustness of GraphBEV to spatial misalignment is taken from other work, not demonstrated in this paper. A single experiment on a small test set cannot support a strong 'proof' of structural difference; a controlled experiment or a systematic comparison is needed.","section":"Section 1, contribution (b); Table 1"},{"comment":"The relationship between the training-time corruption and the test-time misalignment is underspecified. The random variable α is said to replace the current LiDAR frame with a previous frame, but its distribution, the range of lag steps it can sample (e.g., one-frame vs. multi-frame lag), and its value in the reported experiments are never given. Evaluation only tests a fixed one-frame lag. If α rarely samples a one-frame lag or if training uses variable lag lengths, the model's success on the fixed one-frame test could be coincidental. The authors should specify α, perform sensitivity analysis over lag durations, and demonstrate that the learned prediction generalizes across lags.","section":"Section 3.1 and Section 4"}],"minor_comments":[{"comment":"The paper states that mAP is the primary evaluation metric but reports only AP for four object classes. Please also report the overall mAP and NDS on the nuScenes validation set to enable direct comparison with other methods.","section":"Section 4"},{"comment":"The notation Fp, Fo, Fc, and Ff is introduced without fully explaining the feature dimensions (e.g., 336 = 256 + 80) and the computation of the 'deform weights' in the offset convolution module. Please clarify the exact shapes and operations.","section":"Section 3.3"},{"comment":"The method name is written inconsistently as 'Timealign', 'TimeAlign', and in the conclusion as 'Timelign'. Please standardize the spelling.","section":"Throughout"},{"comment":"Reference [29] (Goodnough et al., 'Transfusion medicine—blood transfusion') appears to be unrelated to the topic of 3D object detection and is likely an incorrect citation. Please correct or remove it.","section":"References"},{"comment":"Figures 1, 2, and 3 are referenced in the text but are not embedded in the manuscript; the architecture and modules are difficult to understand without them. Please ensure all figures are included and clearly labeled.","section":"Figures"},{"comment":"The paper says the pre-trained GraphBEV checkpoint is loaded, but it does not specify which checkpoint (e.g., trained with K=8 and sweeps=1) or how the newly initialized modules are trained on top of it. Please provide these details.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant and under-explored problem, and the proposed method is architecturally reasonable. However, the central empirical claim is currently undermined by a confounded comparison (extra LiDAR sweeps in TimeAlign) and very weak evaluation statistics. The 'proof' of a distinction between temporal and spatial misalignment is also overstated. I would encourage the authors to add the missing baselines and ablations, report full validation-set results with error bars, and temper the claims accordingly. The paper is at an early stage and would benefit from substantial revision before publication in a serious venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper tackles a real, under-studied problem: LiDAR frames arriving late relative to cameras in BEV fusion, and proposes a TimeAlign module (Swin-LSTM prediction + deformable combination) on top of GraphBEV. Under a one-frame LiDAR lag on nuScenes they report AP-Car recovery from 0.595 to 0.656. That is worth a look, but the comparison as run does not support the claim that their module is the reason.\n\nThe confound is plain: GraphBEV is evaluated with sweeps=1, so under lag it sees only the delayed T-1 sweep. TimeAlign is fed T-3..T plus the observed frame. The 0.061 AP-Car difference could simply be temporal aggregation, which the paper itself notes is known to help (BEVFusion4D, BEVDet4D). There is no ablation that holds the input history fixed and turns the TimeAlign module off. Without that, the central contribution is not isolated. The stress-test note got this right.\n\nOther soft spots: the 'proof' that temporal misalignment is distinct from spatial is one experiment and a plausibility argument; that is a hypothesis, not a proof. Evaluation is on a small mini test set with no error bars or ablations, and no comparison with existing temporal fusion baselines. The clean-data regression (0.789 vs 0.823) is acknowledged but not probed. No code is provided, so the training corruption (random α frame replacement) is currently the only description of the lag model.\n\nCredit where due: the problem framing is useful and the solution is a reasonable first attempt. The random-drop training procedure is a sensible way to simulate lag. The authors also cite the relevant temporal fusion work and are honest about their limitation. The paper is coherent on its own terms.\n\nFor a reader: this is a workshop-level contribution in its current state, pointing at a gap that deserves a more careful study. I would not cite the numerical result yet.\n\nRecommendation: send it to peer review, but expect major revision. The reviewers should ask for (1) a matched-input baseline that aggregates the same historical sweeps without prediction, (2) ablations that remove the LSTM and combination parts, and (3) at least one comparison with BEVFusion4D or similar under the same lag. If the gain survives that, the paper would make a real point.","headline":"TimeAlign's reported gain over GraphBEV under LiDAR lag is likely just temporal aggregation; the paper needs matched baselines and ablations before the claim holds.","tokens_in":9524,"tokens_out":3687,"would_cite":false,"duration_ms":32283,"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":"TimeAlign, built on GraphBEV, predicts delayed LiDAR BEV features from three historical sweeps and uses camera features to fuse prediction with observation, improving one-frame-lag car AP from 0.595 to 0.656.","keywords":["time misalignment","multi-modal 3D object detection","LiDAR-camera fusion","bird's eye view","temporal feature prediction","autonomous driving","robustness"],"falsifier":"Test TimeAlign on real LiDAR delay traces or on a fixed two-frame lag, and compare against the reported one-frame synthetic lag: if the improvement over GraphBEV vanishes or reverses, the conclusion that the module solves time misalignment rather than overfitting the random-drop training distribution is unsupported.","tokens_in":8544,"feed_emoji":"🚗","tokens_out":5925,"duration_ms":57838,"temperature":0.7,"pith_summary":"This paper argues that temporal misalignment between LiDAR and camera frames, specifically one-frame LiDAR lag, is a distinct failure mode that spatial calibration methods cannot fix, and that it can be mitigated by predicting the delayed LiDAR feature from history. The authors build TimeAlign on the GraphBEV detector: a Swin-LSTM recurrent module forecasts the current LiDAR BEV feature from the previous three frames, and a camera-guided dual-transformer combines that prediction with the actual observed feature. On nuScenes, TimeAlign raises car average precision under one-frame LiDAR lag from 0.595 (GraphBEV baseline) to 0.656, while accepting a smaller drop on perfectly synchronized data (0.789 vs 0.823 for cars). The result matters because real autonomous-driving systems experience LiDAR transfer delays, and perception modules trained on synchronized benchmarks do not currently handle them.","feed_headline":"TimeAlign recovers 3D detection when LiDAR lags the camera","feed_subtitle":"Predicting delayed LiDAR BEV features lifts car AP from 0.595 to 0.656 under a one-frame lag.","key_machinery":"The load-bearing mechanism is a two-part TimeAlign module inserted into GraphBEV's feature fusion. First, a Swin-LSTM recurrent network, an LSTM whose state updates use SwinTransformer-style local-window attention, consumes the LiDAR BEV features of frames T-3 to T-1 and predicts the feature at T, trained with an MSE prediction loss against the true T feature. Second, a camera-guided combination layer concatenates the camera BEV feature with both the predicted and observed LiDAR features, computes learnable offset parameters, and uses deformable convolution to re-align each candidate before fusing them, so the network can weight the stale observation against the forecast depending on how trustworthy the current LiDAR input is.","core_discovery":"The central claim is that LiDAR data lag, simulated by randomly replacing the current LiDAR frame with a previous frame during training, degrades multi-modal detection in a way that space-alignment modules like GraphBEV's cannot absorb, because object and ego motion during the lag is not an affine transformation. TimeAlign's prediction-and-combination design uses three historical LiDAR sweeps to generate a predicted current BEV feature, then lets the camera BEV feature arbitrate between prediction and observation through learnable offsets and deformable convolution. On a one-frame lag the method improves GraphBEV's car AP from 0.595 to 0.656, truck AP from 0.477 to 0.513, bus AP from 0.306 to 0.769, and pedestrian AP from 0.613 to 0.759, at the cost of slightly lower AP on synchronized inputs.","pith_inferences":["The random frame-drop training can be read as a temporal augmentation, so a natural untested extension is camera-side delays or lags of more than one frame.","Real recorded delay traces would be the decisive test of whether the learned behavior transfers beyond the synthetic corruption model.","The camera-guided combination is conceptually a learned Kalman gain; a lighter alternative would estimate per-location observation confidence directly instead of running a recurrent predictor.","On the mini test, bus AP swings from 0.306 to 0.769; whether that holds on the full validation set is a check the paper does not report."],"forward_implications":["Under a one-frame LiDAR lag, TimeAlign improves car, truck, bus, and pedestrian AP over GraphBEV on the nuScenes mini test, with car AP rising from 0.595 to 0.656.","On synchronized inputs, TimeAlign retains most of the baseline performance (car AP 0.789 vs 0.823), indicating that the extra prediction branch does not catastrophically interfere with normal operation.","Because the training procedure injects random frame replacements, the architecture is designed to handle variable lag rather than only the fixed one-frame offset used in evaluation.","The comparison supports the paper's claim that temporal misalignment is structurally different from spatial misalignment and cannot be repaired by affine-style feature re-alignment alone."],"supporting_citations":[{"why":"Supplies the GraphBEV baseline detector and the space-alignment module that the paper shows is insufficient for temporal lag.","marker":"[20]"},{"why":"Provides the synchronized multi-sensor nuScenes sequences and detection labels used for training and the mini-test evaluation.","marker":"[42]"},{"why":"Supplies the Swin-LSTM recurrence that predicts the delayed LiDAR BEV feature.","marker":"[27]"},{"why":"Provides the base LiDAR-camera BEV fusion structure that TimeAlign inherits and extends.","marker":"[9]"},{"why":"Motivates the LiDAR-stuck corruption scenario and the robustness evaluation perspective.","marker":"[18]"},{"why":"Provides the spatio-temporal predictive-learning implementation from which the time alignment module is developed.","marker":"[44]"}],"fun_headline_variants":["TimeAlign predicts LiDAR to fix lag, boosts AP by 0.16","LiDAR delayed? TimeAlign predicts future features for 3D detection","From 0.595 to 0.656 car AP: TimeAlign beats LiDAR lag","TimeAlign: predicting BEV features restores detection under time lag","Handling LiDAR lag: TimeAlign lifts car AP to 0.656"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on treating random frame-drop replacements during training as a faithful stand-in for real LiDAR transfer delays; if actual delays follow a different pattern, the learned corrections may not help on the road.","fun_headline_variants_meta":{"raw":{"variants":["TimeAlign predicts LiDAR to fix lag, boosts AP by 0.16","LiDAR delayed? TimeAlign predicts future features for 3D detection","From 0.595 to 0.656 car AP: TimeAlign beats LiDAR lag","TimeAlign: predicting BEV features restores detection under time lag","Handling LiDAR lag: TimeAlign lifts car AP to 0.656"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000244,"raw_usage":{"total_tokens":1481,"prompt_tokens":845,"completion_tokens":636,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":532}},"tokens_in":461,"tokens_out":636,"duration_ms":6819,"temperature":1.0,"reasoning_tokens":532,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:24:48.923671+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Test TimeAlign on real LiDAR delay traces or on a fixed two-frame lag, and compare against the reported one-frame synthetic lag: if the improvement over GraphBEV vanishes or reverses, the conclusion that the module solves time misalignment rather than overfitting the random-drop training distribution is unsupported.","supporting_citations":[{"cited_title":"Swinlstm: Improving spatiotemporal prediction accuracy using swin transformer and lstm","cited_arxiv_id":null,"evidence_quote":"Supplies the Swin-LSTM recurrence that predicts the delayed LiDAR BEV feature."},{"cited_title":"Benchmarking the robustness of lidar-camera fusion for 3d object detection","cited_arxiv_id":null,"evidence_quote":"Motivates the LiDAR-stuck corruption scenario and the robustness evaluation perspective."},{"cited_title":"Openstl: A comprehensive benchmark of spatio-temporal predictive learning","cited_arxiv_id":null,"evidence_quote":"Provides the spatio-temporal predictive-learning implementation from which the time alignment module is developed."}],"review_version":1}