{"id":"d5f84782-cfb8-4fbb-a7f5-a2175b127576","arxiv_id":"2412.08120","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Dense depth maps can be recovered from a focus-sweep event stream via a voxelized event focal stack and a U-Net, outperforming an image-domain depth-from-defocus baseline on synthetic tests and showing qualitative promise in real low-light scenes.","lead":"This paper turns the stream of events from a sweeping-focus event camera into a dense depth map, using a neural network trained on simulated and real data. The pitch is robustness: depth estimation that keeps working in low light, where ordinary cameras see almost nothing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world claim rests on unaligned pseudo-depth: Depth Anything's relative/affine-invariant output is used as metric inverse-depth ground truth (Sec. 3.2, Eq. 1), with no described scale/shift alignment and no measured-depth evaluation, so real-data 'usability' is unsupported.","rationale":"The central claim has two parts: (i) synthetic event focal stacks trained a U-Net that beats an image-domain depth-from-defocus baseline, and (ii) the method transfers to real events and works in low light. Part (i) is supported by Table 5, though bin size and event simulator are selected on the same 20-scene evaluation set and no code is released. Part (ii) is what makes the contribution distinctive, and its support is Sec. 5/Fig. 7: qualitative depth maps on 10 held-out real scenes, with pseudo-dense depth from Depth Anything used as Dgt for fine-tuning (Sec. 3.2/3.3). Depth Anything's public model outputs relative/affine-invariant depth, not metric inverse depth; using it directly as Dgt in an MSE loss without alignment introduces a scale/shift ambiguity exactly of the kind the reader flagged. If the pseudo labels are scale-ambiguous, the fine-tuned network's inverse-depth output has no calibrated meaning, and qualitative maps can look plausible while being wrong in metric terms. The low-light claim is especially exposed because pseudo-labels come from well-lit images and no quantitative real metric is reported. The missing comparison with the prior event-focal-stack paper [4] is relevant for positioning but does not threaten the internal validity of the synthetic experiment. The strongest check is therefore to quantify the real-data branch against measured depth, with and without alignment of Depth Anything. This concern is the same one the reader identified, so the conditional verdict remains unchanged.","tokens_in":9233,"tokens_out":6672,"duration_ms":75519,"concrete_test":"Capture sensor-measured depth (e.g., LiDAR or structured light) for the 10 real test scenes under both lighting conditions. Fine-tune two models: one with the paper's raw pseudo-Dgt from Depth Anything, and one after per-scene scale/shift alignment of the pseudo-Dgt to sparse measured depths. Evaluate both on the measured depth with MAE/RMSE in 1/m. If the unaligned model is materially worse than the aligned model, or if neither achieves plausible metric accuracy, the real-data claim is not supported by the current protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The fragile step is the real-data supervision. In Sec. 3.2 the network is trained with MSE (Eq. 1) against Dgt, and for real data the paper states 'we utilize pseudo Dgt generated by Depth Anything.' Depth Anything's standard output is relative/affine-invariant depth, not metric inverse depth in 1/m. The paper does not describe any per-scene scale/shift alignment, normalization to the synthetic inverse-depth metric, or filtering of unreliable pseudo labels. Since fine-tuning (Sec. 3.3) is the designated bridge across the synthetic-real domain gap, a mismatch in the target quantity means the model may learn an arbitrary scale or offset from scene to scene. The real evaluation (Sec. 5, Fig. 7) is qualitative only: the 10 held-out scenes have no sensor-measured depth, and the claim that the method 'remains usable in low light' is a visual judgment. This does not invalidate the synthetic comparison, but it leaves the real-data component of the central claim without quantitative support and with a plausible training-signal confound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a method for dense depth estimation from an event stream produced by sweeping the focal plane of a lens attached to an event camera. Events are voxelized into a B=5 'event focal stack' and decoded by a U-Net into a dense inverse depth map. Training uses Blender-rendered focal stacks passed through ESIM or DVS-Voltmeter event simulators, with metric depth ground truth; real-world adaptation is done by fine-tuning on a small real dataset with pseudo ground truth from Depth Anything, after correcting for lens breathing via homographies. The method is compared against a depth-from-defocus baseline (Focus on Defocus) on 20 synthetic test scenes and on 10 real scenes (qualitatively). Quantitative synthetic results show MAE 0.0762 (DVS-Voltmeter) vs 0.1606 for the baseline, with ablations of bin size and polarity integration.","tokens_in":9460,"tokens_out":7454,"duration_ms":71881,"significance":"The idea of using a focus-sweep event stream as a dense depth cue is novel and well-motivated; the event focal stack representation is a natural extension of image focal stacks, and the synthetic-to-real fine-tuning pipeline with lens breathing correction is a practical contribution that goes beyond simple naive domain transfer. The synthetic evaluation is quantitative, uses 20 test scenes, averages over 10 training runs, and clearly separates the proposed configuration from the image-domain baseline. However, the real-world component, which is the abstract's second half of the central claim, is supported only by qualitative depth maps and relies on unaligned pseudo-labels, so the claim of 'superior performance on real datasets' is not yet established.","major_comments":[{"comment":"The abstract claims superior performance over the image-domain baseline on real datasets, but the real-world evaluation is entirely qualitative. No MAE/RMSE or any quantitative metric is reported for the 10 held-out real scenes, and the statement that the method 'remains usable in low light' is a visual judgment. Since the real-data comparison is a load-bearing part of the central claim, please add quantitative evaluation on the real test scenes, ideally against sensor-measured depth (e.g., LiDAR or structured light) or at least against aligned pseudo-labels with clearly stated limitations.","section":"Sec. 5 (Real dataset) and Fig. 7"},{"comment":"The real-data fine-tuning and evaluation use pseudo ground truth Dgt generated by Depth Anything, which outputs affine-invariant (scale- and shift-ambiguous) depth, yet Eq. (1) is an MSE loss in metric inverse depth (1/m). No per-scene scale/shift alignment, normalization to the synthetic inverse-depth range, or filtering of unreliable pseudo-labels is described. This means the fine-tuning target does not match the training target, and the network can internalize an arbitrary scale/offset per scene, confounding both the fine-tuned model and any qualitative comparison. Please describe the alignment procedure, or report results with a properly aligned pseudo-label baseline, or use sensor-measured ground truth.","section":"Sec. 3.2, Eq. (1) and Sec. 4.1 (Real Dataset)"}],"minor_comments":[{"comment":"The paper states that all evaluations were repeated 10 times, but only point estimates are reported. Please add means and standard deviations (or confidence intervals) so the reader can judge the significance of differences, especially for the close B=5 vs B=10 values in Table 2.","section":"Sec. 4.4 and Tables 1-5"},{"comment":"The choice of B=5 for DVS-Voltmeter is inconsistent with the quantitative results: Table 2 gives B=10 the lowest MAE (0.0760) and RMSE (0.1006), while B=5 is second (0.0762, 0.1022). The justification 'plausibly lower errors with B=5' refers to a qualitative figure. Please either adopt the quantitatively best configuration for the main comparison or provide a non-qualitative rationale (e.g., compactness) for choosing B=5, and state how this affects the reported results.","section":"Sec. 4.5 and Table 2"},{"comment":"It is unclear whether the network output is constrained to be non-negative (as inverse depth should be) and how pixels with no events or invalid ground-truth depth are handled in Eq. (1). Please specify the output activation and the masking/weighting of invalid pixels.","section":"Sec. 3.2"},{"comment":"The claim that fine-tuning 'mitigates both the reversed depth and holes' would be strengthened by reporting at least per-scene quantitative errors on the real data after aligning the pseudo-labelled ground truth, rather than relying solely on visual inspection.","section":"Sec. 5 and Fig. 7"},{"comment":"The citation '[27]' is referred to in the text as 'Alex et al.'; for consistency with the reference list it should be 'Zhu et al.' Also, the figure captions contain inline code-like references such as '(Sec 3.1)' that should be formatted consistently.","section":"Figure 1 and Sec. 3.1"}],"recommendation":"major_revision","confidential_remarks":"The main weakness is the real-data evaluation. If the authors can provide a quantitative real-world evaluation with aligned ground truth (or at least transparently report the alignment procedure and errors against aligned pseudo-labels), the contribution would meet the bar for publication. The synthetic comparison is solid but narrow (one baseline); a comparison with the sparse event-depth methods cited in Sec. 2.2 would strengthen the novelty claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a plausible first step on dense depth from event focal stacks, and the strongest part is the synthetic pipeline. Prior work on event focal stacks estimated focus distance or sparse depth; this paper outputs dense inverse depth from a 5-bin voxelized event focal stack through a U-Net, and it does so with a sensible Blender-plus-event-simulator training recipe and a lens-breathing correction for real captures. The synthetic comparison is honest and reasonably careful: 20 test scenes, 10 training runs, and a clear margin over a properly trained image-domain DfD baseline (Focus on Defocus), with MAE 0.0762 vs 0.1606 for DVS-Voltmeter. That advantage is large enough to be credible, not just seed noise.\n\nThe soft spots are real but concentrated in the real-data claim. Fine-tuning uses Depth Anything pseudo-depth as metric inverse-depth ground truth, yet Depth Anything's output is not metric inverse depth in 1/m, and the paper never describes scale/shift alignment, normalization, or filtering of unreliable pseudo-labels. So the fine-tuned model may be learning scene-dependent scale or offset rather than true depth. The 10 real test scenes are evaluated only qualitatively; there is no sensor-measured depth and no MAE/RMSE on real data. That does not invalidate the synthetic comparison, but it leaves the abstract's 'superior performance on real datasets' unsupported. A smaller issue: the B=5 bin choice is selected on the evaluation set in Tables 1 and 2, which mildly inflates the headline number; not fatal, but it should be reported with variance or a held-out split. No code or data are released, and the relationship to the closest event-focal-stack work [4] is not fully resolved—they cite it as sparse depth but do not compare against it.\n\nThis paper deserves a serious referee. The real-data section needs metric ground truth and a fix for the Depth Anything scale problem; the synthetic study is worth publishing on its own. I would bring it to a reading group and would cite it as the first dense event-focal-stack depth method, with a caveat on the real evaluation.","headline":"A credible dense-depth-from-event-focal-stack pipeline with a strong synthetic comparison, but the real-data claim rests on unaligned pseudo-depth and qualitative evaluation.","tokens_in":10008,"tokens_out":2048,"would_cite":true,"duration_ms":22646,"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":"Dense depth maps from an event camera focus sweep","keywords":["event camera","dense depth estimation","depth from defocus","event focal stack","focal sweep","event simulation","low-light vision","encoder-decoder network"],"falsifier":"Record a focus-sweep event stream in a scene with independent measured depth (for example, a LiDAR or structured-light scan) in low-light conditions, run the proposed network, and compare inverse depth predictions against the measured values; if the predictions do not beat the image-domain baseline, or if the MAE against true depth is no better than chance, the central claim fails.","tokens_in":8995,"feed_emoji":"⚡","tokens_out":4910,"duration_ms":46547,"temperature":0.7,"pith_summary":"This paper claims that a dense inverse depth map can be predicted from the event stream produced while an event camera's lens sweeps its focal plane. It turns the stream into a five-bin event focal stack and decodes it with a U-Net trained on simulated events. On synthetic scenes the method roughly halves the error of an image-domain depth-from-defocus baseline, and on real scenes it still produces usable depth in low light after fine-tuning on a small set of real events. The practical point is a depth sensor that works where ordinary cameras fail, using the event camera's high dynamic range and low power draw.","feed_headline":"Event-camera focus sweep predicts dense depth maps","feed_subtitle":"It beats image-based depth-from-defocus and still works in dark scenes where ordinary cameras see nothing.","key_machinery":"The load-bearing object is the event focal stack: a $B=5$ channel voxel grid of time-weighted event counts, built by normalizing event timestamps into bins and distributing each event's polarity between the two closest bins. This compact representation lets a U-Net-style encoder-decoder with skip connections see the temporal order of defocus-induced events and produce a dense depth map with mean squared error loss against inverse depth. Lens breathing from real focus sweeps is corrected by homography warps computed from a defocus-robust circular checkerboard. Synthetic training data comes from rendered focal stacks of random 3D scenes converted to events by two event simulators, and the real-data step fine-tunes on fifty scenes with pseudo ground truth from a pre-trained monocular depth estimator.","core_discovery":"The central claim is that the spatial and temporal structure of a focus-sweep event stream encodes depth densely enough to regress an inverse depth map at every pixel, not just at event locations. The paper defends this by constructing event focal stacks from events and training a convolutional encoder-decoder to map the stack directly to inverse depth, using synthetic focal stacks rendered from arbitrary 3D scenes and converted to events through event simulators. The reported result is that this event-based method outperforms a comparable image-domain depth-from-defocus network (MAE 0.0762 vs 0.1606 in inverse depth on synthetic data), and that after fine-tuning on real events it gives plausible qualitative depth in office and low-light scenes.","pith_inferences":["One testable extension is to add sensor-measured ground truth (stereo or LiDAR) to the real benchmark; the paper's real-data evaluation currently rests on pseudo depth from a monocular estimator, so a true depth comparison would settle the real-world claim.","The authors observe that simulators produce few negative events; building a simulator with better noise and polarity statistics could remove much of the need for fine-tuning.","Applying the same event focal stack to dynamic scenes would require aligning events on moving objects to a common sweep; the paper leaves that as a limitation, so a natural follow-up is a recurrent or motion-compensated variant."],"forward_implications":["Depth sensing can continue under low light or rapid motion, where frame-based cameras lose the signal.","The dense-depth task can be trained without real depth sensors, since synthetic focal stacks and event simulation provide labeled data at scale.","A focus-tunable lens, if it can sweep faster than a mechanical lens, should feed the same network with more events per unit time and improve accuracy.","The success of the event focal stack suggests treating a focus sweep as a time-coded volume rather than a set of images, which may also help other event-based vision tasks."],"supporting_citations":[{"why":"Supplies the time-weighted event voxelization and the U-Net architecture that the paper adapts from gray-scale video reconstruction to depth.","marker":"[12]"},{"why":"The image-domain depth-from-defocus framework used as the baseline that the proposed method claims to outperform.","marker":"[16]"},{"why":"One of the two event simulators used to generate synthetic training events from rendered focal stacks.","marker":"[20]"},{"why":"The second event simulator, based on sensor voltage dynamics, used for synthetic training and for studying the simulator-to-real gap.","marker":"[22]"},{"why":"Provides pseudo ground truth depth for real-world images, used for fine-tuning and for qualitative real-data evaluation.","marker":"[25]"},{"why":"Defines the voxelization scheme the event focal stack construction follows.","marker":"[27]"}],"fun_headline_variants":["Event focal stack yields dense depth in the dark","Focus-sweep events resolve depth at every pixel","Event camera sees depth where images fail","Depth from events without visible light","Sweep focus, get dense map: event camera depth"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The real-world evaluation and fine-tuning use pseudo depth from a pre-trained monocular depth estimator applied to well-lit images, not sensor-measured depth; if that pseudo depth is biased, the fine-tuned model inherits the bias and the real-data claim is not demonstrated.","fun_headline_variants_meta":{"raw":{"variants":["Event focal stack yields dense depth in the dark","Focus-sweep events resolve depth at every pixel","Event camera sees depth where images fail","Depth from events without visible light","Sweep focus, get dense map: event camera depth"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1355,"prompt_tokens":782,"completion_tokens":573,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":398,"completion_tokens_details":{"reasoning_tokens":503}},"tokens_in":398,"tokens_out":573,"duration_ms":5699,"temperature":1.0,"reasoning_tokens":503,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:11:56.632565+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a focus-sweep event stream in a scene with independent measured depth (for example, a LiDAR or structured-light scan) in low-light conditions, run the proposed network, and compare inverse depth predictions against the measured values; if the predictions do not beat the image-domain baseline, or if the MAE against true depth is no better than chance, the central claim fails.","supporting_citations":[{"cited_title":"High speed and high dynamic range video with an event camera","cited_arxiv_id":null,"evidence_quote":"Supplies the time-weighted event voxelization and the U-Net architecture that the paper adapts from gray-scale video reconstruction to depth."},{"cited_title":"Fo- cus on defocus: Bridging the synthetic to real domain gap for depth estimation","cited_arxiv_id":null,"evidence_quote":"The image-domain depth-from-defocus framework used as the baseline that the proposed method claims to outperform."},{"cited_title":"ESIM: an open event camera simulator","cited_arxiv_id":null,"evidence_quote":"One of the two event simulators used to generate synthetic training events from rendered focal stacks."},{"cited_title":"Dvs- voltmeter: Stochastic process-based event simulator for dy- namic vision sensors","cited_arxiv_id":null,"evidence_quote":"The second event simulator, based on sensor voltage dynamics, used for synthetic training and for studying the simulator-to-real gap."},{"cited_title":"Depth anything: Unleashing the power of large-scale unlabeled data","cited_arxiv_id":null,"evidence_quote":"Provides pseudo ground truth depth for real-world images, used for fine-tuning and for qualitative real-data evaluation."},{"cited_title":"Unsupervised event-based optical flow using mo- tion compensation","cited_arxiv_id":null,"evidence_quote":"Defines the voxelization scheme the event focal stack construction follows."}],"review_version":1}