{"id":"d038f9df-b13e-4c05-8bcc-70585b30ac1c","arxiv_id":"2505.02593","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DELTA fuses event camera and LiDAR data with transformer attention and recurrent memory to estimate dense depth maps, reporting large close-range error reductions over prior work.","lead":"This paper presents DELTA, a neural network that combines event camera and LiDAR data to produce dense depth maps. It reports up to four times lower error at close ranges than the previous best method, on several driving datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 4x close-range gain is measured against ALED restricted to D_bf; if ALED's intended event-level output is better, the SOTA claim is overstated.","rationale":"The paper is a solid engineering contribution: the ablations (Tab. 6) show that the propagation memory, central memory, and cross-attention module all contribute, and the code and trained models are promised. My concern is not about the architecture but about the yardstick used for the headline. The abstract's 'up to four times' is anchored to Tab. 2, where the only LiDAR-fusion baseline on SLED is ALED, and it is evaluated on D_bf only. ALED's original design produces two depths per window and uses events to assign depths; if the event-level output is the intended SOTA result, then restricting ALED to D_bf may not represent the previous SOTA. The paper's own Sec. 3.1 argues the two-depth formulation is ill-posed, but that argument concerns the formulation, not the measured accuracy of the baseline. Without ALED's full-output numbers on the same SLED splits, the claimed 4x reduction could be an artifact of the comparison protocol. This reinforces the reader's CONDITIONAL verdict: the central claim is plausible but not yet independently verified. MVSEC ground-truth errors and M3ED split redefinition are secondary; they affect real-world generalization but not the headline close-range number on synthetic SLED.","tokens_in":17791,"tokens_out":11037,"duration_ms":130192,"concrete_test":"Run the released ALED code on the SLED test sequences using its full event-level output (assigning each event its d_bf or d_af according to event timestamps) and recompute the 10m mean depth error. Compare this value to DELTA's 0.54m on Town03. If the event-level error is substantially below 2.01m, the claimed 4x reduction is overstated; if it remains near 2.01m, the D_bf-only comparison is fair and the headline claim holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim rests on the SLED comparison in Tab. 2, where DELTA reduces the Town03 10m mean error from 2.01m to 0.54m. The comparison target is ALED evaluated only on its D_bf depth maps (Sec. 4.2). ALED was designed to output two depth maps per temporal window and to use event timestamps to assign each event its d_bf or d_af; its published SOTA results may have been obtained with this full event-level output, not with the D_bf map alone. The paper justifies the D_bf-only comparison by noting that DELTA outputs one depth per window, but fairness requires comparing against ALED's best or intended output. If ALED's event-level depth is more accurate at close range, the 4x reduction is inflated and the 'new state of the art' claim is not established. The manuscript provides no numbers for ALED's full output on the same SLED splits, so the comparison target is unverified. This is a concrete, testable protocol issue rather than a disagreement with the architecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DELTA, a recurrent attention-based network that fuses projected LiDAR point clouds with temporal windows of event data to produce dense depth maps. The architecture uses self- and cross-attention in a U-Net-like structure, with a propagation memory for temporally propagating LiDAR data using events and a GRU-based central memory for temporal smoothing. The method is evaluated on three datasets (SLED, MVSEC, and a redefined subset of M3ED) and compared with prior event-and-LiDAR depth estimation methods. The main claims are that DELTA sets a new state of the art and that it reduces depth errors up to four times at close ranges compared to the previous state of the art, based on SLED Town03 results against ALED.","tokens_in":17996,"tokens_out":3471,"duration_ms":45161,"significance":"The core architectural idea is well motivated: attention-based cross-modal fusion with two memory mechanisms is a sensible and timely contribution, and the ablation study in Sec. 4.6 provides evidence that each component matters. The paper is also commendable for releasing source code and trained models, for performing ablations beyond the main comparison, and for reporting computational complexity in the supplementary material. If the reported gains hold under a fair evaluation protocol, the SLED results represent a substantial improvement in close-range depth estimation. However, the headline state-of-the-art claim and the factor-of-four improvement are conditional on a comparison protocol whose fairness is not fully established: the primary baseline is evaluated only on one of its two output depth maps, and no variance estimates are provided. The acknowledged ground-truth limitations of MVSEC and M3ED further temper the strength of the cross-dataset conclusions.","major_comments":[{"comment":"The paper states that 'for fairness of evaluation, comparisons with ALED [3] are only made on its D_bf depth maps,' but ALED was designed to output two depth maps per temporal window and to use event timestamps to assign each event its d_bf or d_af. The headline claim of a roughly four-fold error reduction at 10m on SLED Town03 (2.01m vs. 0.54m) is based entirely on this D_bf-only comparison. If ALED's intended event-level output is more accurate, then both the factor-of-four improvement and the 'new state of the art' claim are overstated. Please report ALED's full output (or its best available output) on the same SLED and MVSEC splits, or provide a principled argument for why D_bf-only is the correct comparison target.","section":"Sec. 4.2, Tables 2 and 3"},{"comment":"The M3ED dataset is subsampled and the train/val/test splits are redefined because LiDAR data is not provided for the original test set. The ALED results in Tab. 5 are stated to have been 'computed for this article,' which is appropriate, but the paper must confirm that ALED and DELTA were evaluated on the exact same redefined split and with the same subsampling and preprocessing. Without this confirmation, the M3ED comparison is not reproducible and the improvements reported there may be split-dependent. The exact split definition should be provided in the supplementary material.","section":"Sec. 4.1, Tab. 1, Sec. 4.5"},{"comment":"All quantitative results appear to come from a single training run. Several conclusions rely on small differences between methods, for example in Tab. 5 where DELTA_M3 and ALED_M3 differ by only 0.005 to 0.01 in the delta_1/delta_2 metrics, and in Tab. 6 where the central-memory ablation differs by 0.09m at the 200m cutoff. Without error bars, multiple seeds, or statistical significance tests, the robustness of these gains cannot be assessed. Please provide mean and standard deviation over at least three training seeds for the main tables, or explicitly justify why a single run is sufficient.","section":"Secs. 4.3-4.6, Tables 2-7"}],"minor_comments":[{"comment":"The text refers to the SLED test sequences as 'Town01' and 'Town03', while the figure captions use 'Town01_08' and 'Town03_19'; please be consistent.","section":"Sec. 4.3, Fig. 3"},{"comment":"The patch size P and dimensionality D are given, but the relationship between the number of patches N and the input resolution is not stated. This matters for interpreting the propagation memory size of 128, which is much smaller than the typical number of patches for a 512x512 input. A short sentence explaining how N is computed would improve clarity.","section":"Sec. 4.2"},{"comment":"The table uses 'FLOPS' where the standard unit is 'FLOPs'; also, inference-time measurements should state whether they include GPU warm-up and which PyTorch version was used, to make the comparison with ALED more meaningful.","section":"Supplementary Material, Tab. 8"},{"comment":"The bibliography entry for Cui et al. [10] contains stray page numbers after the year ('2022. 2, 5, 6, 7, 8, 1, 4'); these appear to be citation markers and should be removed.","section":"References"},{"comment":"The paper acknowledges that MVSEC ground truth is 'erroneous when there are moving objects in the scene,' yet MVSEC is still used for quantitative evaluation. This is a known limitation of the dataset, but it would be helpful to state explicitly in the main text whether any filtering of erroneous ground-truth pixels was applied, and to discuss how this could affect the reported numbers.","section":"Sec. 4.4"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the D_bf-only comparison against ALED. If the authors can provide ALED's event-level outputs on the same splits and show that the improvement persists, the paper's central claim would be substantially stronger. Please also ask for error bars or multiple seeds on the main comparisons before final acceptance. The paper is otherwise within the scope of a computer vision journal and the code release is a positive factor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about arXiv:2505.02593. First, the DELTA architecture is genuinely new: self- and cross-attention fusion with a propagation memory and a GRU-based central memory, evaluated on three datasets with ablations. Second, the 'up to four times' close-range improvement over the previous SOTA is real on the SLED Town03 sequence, but it is measured against ALED restricted to its D_bf depth maps, and that choice is doing more work than the paper lets on. The authors justify it by defining the task as one depth per temporal window, but ALED was designed to output two depths per window and assign each event its before/after depth. If ALED's full event-level output is more accurate at close range, the 4x gain and the 'new SOTA' claim are overstated. The paper does not provide ALED's full-output numbers on the same splits. This is a testable protocol issue, not a disagreement with the architecture.\n\nWhat is good: the ablation study is thorough, and it supports the importance of both memories and the cross-attention module. The paper is transparent about known problems with MVSEC ground truth and M3ED sparsity, and it shows failure cases. Code and trained models are promised, which helps reproducibility. The complexity comparison with ALED is useful.\n\nSoft spots beyond the D_bf issue: no error bars or repeated training runs, so the close-range numbers, especially the 4x on a single sequence, could be partly luck. On longer ranges the gains shrink or vanish — at the 200m cutoff on Town01 DELTA's mean error is actually 4.54 vs ALED's 4.53. The M3ED evaluation uses a redefined and subsampled split, and the ground truth there is sparse enough to make the numbers less informative. These are not fatal; the paper acknowledges them, but they cap how strongly you can state SOTA.\n\nWho is this for? People working on event-camera depth estimation or LiDAR densification. It deserves a serious referee. I'd accept it for review, but I'd ask the authors to add error bars, report ALED's full event-level output for the SLED comparison, and soften the SOTA claim to 'on window-level dense depth from events and LiDAR.'","headline":"A credible attention-based fusion architecture with honest limitations, but the '4x better' claim hinges on comparing ALED only on its D_bf output, so the SOTA claim is broader than the evidence.","tokens_in":18528,"tokens_out":2911,"would_cite":true,"duration_ms":35665,"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":"DELTA uses self- and cross-attention to fuse event-camera and LiDAR data, cutting close-range depth error by up to four times.","keywords":["event camera","LiDAR fusion","dense depth estimation","transformer attention","cross-attention","recurrent memory","autonomous driving","depth from events"],"falsifier":"Run DELTA and the previous method on a dataset with dense, synchronized, error-free ground truth, or on MVSEC after correcting its known moving-object depth errors, and compare every output, including the previous method's after-event depth maps. If the 10 m cutoff mean error on SLED Town03 does not stay near 0.54 m versus 2.01 m, the central improvement claim is refuted.","tokens_in":17583,"feed_emoji":"🚗","tokens_out":9480,"duration_ms":89822,"temperature":0.7,"pith_summary":"This paper tries to establish that an attention-based network, DELTA, can fuse asynchronous event-camera data with sparse LiDAR depth to produce dense depth maps, and that this fusion strategy beats the previous state of the art, especially close to the sensor. On the synthetic SLED dataset, the mean depth error at the 10 m cutoff drops from 2.01 m to 0.54 m on the Town03 test sequence, roughly a fourfold reduction. The authors argue the gain comes from replacing convolutional fusion with self- and cross-attention, supported by two memory mechanisms: a propagation memory that updates stale LiDAR with newer events and a central recurrent memory that stabilizes predictions when the platform stops. If the claim holds, event-plus-LiDAR depth estimation becomes substantially more accurate for the close-range driving scenarios where pedestrians and obstacles matter most.","feed_headline":"Close-range depth error drops 4x with event-LiDAR attention","feed_subtitle":"DELTA densifies sparse LiDAR with event streams and beats prior depth estimators, especially near the vehicle.","key_machinery":"The load-bearing machinery is attention with two memories. Event volumes and projected LiDAR point clouds are split into patches and self-attended separately; a central cross-attention module then fuses the two modalities; a propagation memory, queried by the previous LiDAR representation, updates the LiDAR features with the newest events even when no new LiDAR sweep has arrived; and a GRU maintains a central memory state that carries context across time. The propagation memory is what lets fusion happen at the event rate rather than the LiDAR rate, and the ablation study identifies the central cross-attention module as the component whose removal costs the most accuracy.","core_discovery":"The central claim is that dense depth from events and LiDAR is better solved by attention-based fusion than by the convolutional fusion used in the prior state of the art. DELTA tokenizes event volumes and projected LiDAR into patches, applies self-attention within each modality and cross-attention between them, and maintains two memories: a small propagation memory that lets stale LiDAR be updated by newer events, and a gated recurrent unit (GRU) central memory that condenses the past and smooths the output. Evaluated on SLED, MVSEC, and M3ED, the paper reports a new state of the art, with the largest gains at close range, up to about four times lower mean depth error at the 10 m cutoff on SLED. The paper also revises its earlier two-depth-map formulation, arguing that only the before-event depth map per temporal window is needed because the after-event map can be read from the next window.","pith_inferences":["An implication the paper leaves implicit is that the close-range gain should depend on LiDAR density; sparser LiDAR would likely shrink the gap between attention and convolutional fusion, a testable prediction the paper does not run.","Because DSEC was excluded on the grounds that its ground truth is timestamped to RGB frames rather than LiDAR, resynchronizing DSEC depth to LiDAR timestamps would provide a larger real-world benchmark for testing whether the SLED gains transfer.","The revision from two depth maps to one implies after-event depth can be recovered from the next window's before-event map, so a production version could halve the output head's cost without losing information.","The reported failure cases, such as event-volume saturation during fast turns and dark objects missed by the event camera, suggest the event representation, not the attention mechanism, is the next bottleneck; time-ordered or asynchronous event representations are a natural extension."],"forward_implications":["At the 10 m cutoff on the SLED Town03 sequence, mean depth error drops from 2.01 m with the previous best method to 0.54 m with DELTA, roughly a fourfold reduction, with smaller but consistent gains on Town01.","Because the propagation memory updates LiDAR features with each incoming event window, fusion can run at the event stream's temporal resolution instead of waiting for the next LiDAR sweep.","Ablations attribute the improvement to specific components: removing the central cross-attention hurts most, removing the propagation memory adds roughly 0.2–0.4 m of error, and removing the central memory adds up to about 0.3 m at close range.","DELTA remains competitive on real-world MVSEC and M3ED data despite acknowledged errors and sparsity in their ground truth, and pretraining on synthetic SLED transfers reasonably to real sensors.","On high-resolution 1280x720 input, DELTA runs at about 6.3 Hz with roughly 1.79 TFLOPs on a single GPU, with FLOPs and memory usage comparable to the previous convolutional method."],"supporting_citations":[{"why":"Supplies the SLED dataset and the ALED baseline whose before-event depth outputs are the main comparison, as well as the earlier two-depth-map formulation this work revises.","marker":"[3]"},{"why":"Provides the M3ED dataset whose train, validation, and test splits are redefined and subsampled for this evaluation.","marker":"[5]"},{"why":"Provides the gated recurrent unit used to update the central memory for temporal stability.","marker":"[8]"},{"why":"Provides the geometry-based event-and-LiDAR depth method used as a comparison point in the quantitative and qualitative evaluation.","marker":"[10]"},{"why":"Provides the patch-based splitting and fixed positional embedding that let the network apply attention to image-like event and LiDAR data.","marker":"[12]"},{"why":"Provides the U-Net encoder-decoder structure that organizes the two input branches, skip connections, and decoding branch.","marker":"[34]"},{"why":"Provides the convex upsampling modules used in the decoding head to regroup patches into full-resolution depth maps.","marker":"[42]"},{"why":"Provides the multiscale gradient-matching loss used alongside the L1 loss to keep predicted depth maps sharp.","marker":"[44]"},{"why":"Provides the MVSEC dataset and the cutoff-distance evaluation convention used for the real-world comparison.","marker":"[49]"}],"fun_headline_variants":["Event-LiDAR attention fusion slashes close-range depth error 4x","Transformer attention for dense depth from events and LiDAR","Up to 4x lower depth error via event-LiDAR attention fusion","DELTA model sets new SOTA for event-based depth estimation","Cross-attention between events and LiDAR beats prior depth methods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art claim rests on the evaluation protocol: if the imperfect ground truth in MVSEC, the redefined M3ED splits, or the choice to compare the previous method only on its before-event depth outputs unfairly favors DELTA, the fourfold close-range improvement is overstated.","fun_headline_variants_meta":{"raw":{"variants":["Event-LiDAR attention fusion slashes close-range depth error 4x","Transformer attention for dense depth from events and LiDAR","Up to 4x lower depth error via event-LiDAR attention fusion","DELTA model sets new SOTA for event-based depth estimation","Cross-attention between events and LiDAR beats prior depth methods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000284,"raw_usage":{"total_tokens":1635,"prompt_tokens":863,"completion_tokens":772,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":682}},"tokens_in":479,"tokens_out":772,"duration_ms":8143,"temperature":1.0,"reasoning_tokens":682,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:46:50.589388+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DELTA and the previous method on a dataset with dense, synchronized, error-free ground truth, or on MVSEC after correcting its known moving-object depth errors, and compare every output, including the previous method's after-event depth maps. If the 10 m cutoff mean error on SLED Town03 does not stay near 0.54 m versus 2.01 m, the central improvement claim is refuted.","supporting_citations":[{"cited_title":"Learning to estimate two dense depths from LiDAR and event data","cited_arxiv_id":null,"evidence_quote":"Supplies the SLED dataset and the ALED baseline whose before-event depth outputs are the main comparison, as well as the earlier two-depth-map formulation this work revises."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the M3ED dataset whose train, validation, and test splits are redefined and subsampled for this evaluation."},{"cited_title":"Learning phrase representations using RNN encoder-decoder for statistical machine translation","cited_arxiv_id":null,"evidence_quote":"Provides the gated recurrent unit used to update the central memory for temporal stability."},{"cited_title":"Dense depth-map estimation based on fusion of event camera and sparse LiDAR.IEEE Transactions on Instrumentation and Measurement, 71:1– 11, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the geometry-based event-and-LiDAR depth method used as a comparison point in the quantitative and qualitative evaluation."},{"cited_title":"An image is worth 16x16 words: Transformers for image recognition at scale","cited_arxiv_id":null,"evidence_quote":"Provides the patch-based splitting and fixed positional embedding that let the network apply attention to image-like event and LiDAR data."},{"cited_title":"U- Net: Convolutional networks for biomedical image segmen- tation","cited_arxiv_id":null,"evidence_quote":"Provides the U-Net encoder-decoder structure that organizes the two input branches, skip connections, and decoding branch."},{"cited_title":"RAFT: Recurrent all-pairs field transforms for optical flow","cited_arxiv_id":null,"evidence_quote":"Provides the convex upsampling modules used in the decoding head to regroup patches into full-resolution depth maps."},{"cited_title":"DeMoN: Depth and motion network for learning monocular stereo.CVPR, pages 5622–5631, 2016","cited_arxiv_id":null,"evidence_quote":"Provides the multiscale gradient-matching loss used alongside the L1 loss to keep predicted depth maps sharp."},{"cited_title":"Kumar, and Kostas Daniilidis","cited_arxiv_id":null,"evidence_quote":"Provides the MVSEC dataset and the cutoff-distance evaluation convention used for the real-world comparison."}],"review_version":1}