{"id":"88e320f8-68f2-4dde-bf0d-e112daf0aec4","arxiv_id":"2504.17728","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A one-stage pipeline jointly estimates camera motion, exposure times, and the camera response while reconstructing an HDR 3D scene from blurry auto-exposure videos.","lead":"This paper introduces a method that builds 3D high-dynamic-range scenes from ordinary phone or camera videos shot with auto-exposure, even when the video is blurry and exposure settings change between frames. It could make 3D HDR capture practical for everyday devices and support VR, AR, and relighting applications.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exposure-time identifiability is only supported by rank correlations; the flexible CRF and per-image WB can absorb brightness differences that should constrain exposure times.","rationale":"The reader's weakest assumption correctly identifies the identifiability of exposure times and CRF as the load-bearing premise, and the paper's own evidence is insufficient: Appendix A.1 only reports rank correlations on real data, and the loss's mean-normalization in Eq. 12 removes the absolute brightness constraint. The synthetic dataset, which has ground-truth exposure times, is not used for absolute exposure validation, which would be the natural check. This concern does not invalidate the empirical rendering results, but it means the central physical-identifiability claim is not established. The appropriate verdict remains CONDITIONAL, matching the reader's judgment; my stress-test adds a concrete test that would either confirm or resolve the concern.","tokens_in":17076,"tokens_out":3826,"duration_ms":40987,"concrete_test":"On the synthetic Blender scenes (Factory, Pool, Trolley, Cozyroom), where ground-truth exposure times, CRF, and HDR radiance are known, train Casual3DHDR-random, then: (1) fit a single global scale factor to align estimated exposure times with ground truth and compute per-frame median absolute relative error; (2) render novel views at exposure multipliers 1x, 2x, and 4x and compare against ground-truth renders using an HDR-aware metric such as PU21-PSNR. If median relative exposure error exceeds 20% or the exposure-scaled renders degrade sharply, the method does not recover physical exposure times.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that unknown, varying exposure times are actually recovered, not merely that the network can fit appearance. In Sec. 3.3, Eq. 7-9 compose exposure time, scene radiance H, per-image white balance WB, and per-channel tone-mapping MLPs TM. This composition has a fundamental scale ambiguity: multiplying H by s and dividing every exposure time by s leaves the integral H*dt unchanged, so the product is identifiable only up to a global scale. More seriously, the per-image WB (Eq. 10) and the flexible per-channel MLP for TM can absorb frame-to-frame brightness differences without changing exposure times at all. The loss in Eq. 12 normalizes each image by its mean pixel value before computing L1 and D-SSIM, so the optimization does not constrain the absolute brightness scale. The only evidence for exposure-time recovery is Appendix A.1, which reports Pearson, Spearman, and Kendall correlations on two real sequences. These rank-based metrics are invariant to any monotonic transformation, so they do not show that the estimated times are accurate in any absolute sense. Notably, the paper does not report absolute exposure-time error on the synthetic Blender scenes where ground-truth exposure times are known. If the estimated exposure times and CRF are appearance-fitting artifacts, then the claimed HDR radiance field and exposure editing capability are not physically validated, and the core contribution -- joint physical optimization of exposure times, CRF, and trajectory -- is unproven.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Casual3DHDR, a one-stage method for reconstructing 3D high dynamic range (HDR) scenes from casually captured auto-exposure videos that contain motion blur, brightness variation, and unknown, varying exposure times. The method represents camera motion with a cumulative SE(3) B-spline continuous-time trajectory, models motion blur as an average over virtual sharp frames during the exposure interval, and jointly optimizes the 3D Gaussian splatting scene, the trajectory, per-frame exposure times, and an implicit camera response function (CRF) composed of per-image white balance and per-channel tone-mapping MLPs. The loss combines an exposure-normalized reconstruction term with a D-SSIM term. Experiments on synthetic Blender scenes and real RealSense/Pixel/ScanNet sequences compare novel view synthesis, deblurring, and pose accuracy against HDR-NeRF, HDR-Plenoxels, Gaussian-W, BAD-Gaussians, and gsplat baselines, with ablations on trajectory initialization and on each model component.","tokens_in":17420,"tokens_out":4841,"duration_ms":49814,"significance":"If the central claim holds, the paper offers a practical advance: consumer-grade AE videos, with unknown exposure times and motion blur, would become sufficient input for HDR 3D scene reconstruction without calibrated exposure settings. The synthetic experiments cover the full blur-and-exposure setting and the proposed components are individually ablated, which is a strength. The dataset release is also a useful contribution. The NVS evaluation on held-out views against external baselines is not circular. However, the validation of the load-bearing exposure-time recovery is weak: only rank correlations on two real sequences are provided, and the flexible CRF/white-balance model may absorb the brightness signal that should constrain exposure times. I therefore regard the result as promising but not yet fully established.","major_comments":[{"comment":"The paper does not establish that exposure times are identifiable from the proposed loss. The forward model in Eqs. (7)–(9) contains per-image white-balance gains (Eq. 10) and flexible per-channel tone-mapping MLPs, while the loss in Eq. (12) removes the absolute brightness scale by normalizing each image by its mean pixel value. There is a global scale ambiguity between the HDR radiance H and Δt, and more seriously, the white-balance and CRF components can absorb frame-to-frame brightness variations that should constrain Δt. The only quantitative validation, Table 8, reports rank correlations on two real sequences; rank metrics are invariant to monotonic transformations, and the correlations are computed on the same videos used for optimization, so they do not demonstrate absolute or predictive accuracy. Because recovery of unknown, varying exposure times is a central contribution, please report absolute or relative exposure-time error on the synthetic Blender scenes where ground-truth exposure times are known, and provide either an identifiability argument or an ablation that constrains the CRF/white-balance model to isolate the exposure-time signal.","section":"§3.3–3.4 and Appendix A.1"},{"comment":"The real-world NVS evaluation uses only manually selected sharp frames: the text states \"we used 5 to 10 sharp images per sequence for evaluation.\" This protocol does not exercise the severe motion blur that is the motivating condition of the paper, and a method that fails on blurry frames could still score well on this sharp subset. Please evaluate on all captured frames, report sharp-frame and blurry-frame results separately, and specify the exact selection protocol and per-sequence numbers of selected frames.","section":"§4.4 and Tables 2/10"},{"comment":"No error bars or significance tests are reported; all comparisons are single runs despite stochastic components in training, including MCMC-based densification in gsplat and random exposure-time initialization (Sec. 4.2). For claims such as those in Table 5, where ATE differences between methods are fractions of a centimeter, the absence of variance estimates makes the claimed superiority impossible to assess. Please report mean±std over multiple seeds or runs, and where appropriate, paired significance tests.","section":"Tables 1–8"}],"minor_comments":[{"comment":"The sentence \"the results show that we can estimate exposure times :\" is incomplete and should be finished or removed; also, Figure 9 plots estimates that are scaled uniformly, so the figure cannot be read as evidence of absolute accuracy.","section":"Appendix A.1"},{"comment":"Several typos remain, for example \"Sythetic\" (Sec. 4.2), \"server motion blur\" (Sec. 4.3), \"image domian translation\" (Sec. 2.1), and \"Acknowldement\" (Section 6).","section":"§4.2, §4.3, §2.1, §6"},{"comment":"The citation [34] is used for NeRF, but [34] is NeRF-W; the correct NeRF reference appears to be [36], so the citation should be corrected.","section":"§2.3"},{"comment":"Table 9 duplicates the BAD-Gaussians rows of Table 3, and Table 10 repeats the real-dataset rows of Table 2; the supplementary material should refer to the main tables instead of duplicating them.","section":"Appendix A.2/A.3"},{"comment":"The CRF module is described only as per-channel MLPs; please specify the architecture details, including number of layers, hidden size, activation function, and initialization, in the implementation section.","section":"§3.3 and §4.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable fit for a multimedia venue and the core idea is interesting. I do not see grounds for rejection, but the exposure-time identifiability issue is load-bearing and currently supported only by rank correlations; the requested synthetic absolute-error evaluation and identifiability discussion are necessary before the contribution can be considered established. The evaluation protocol for real-world NVS also needs strengthening. I recommend major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nCasual3DHDR is a genuine step forward: it is the first 3DGS-based HDR reconstruction I know that takes casually captured auto-exposure video, with motion blur and unknown exposure times, and jointly optimizes a continuous SE(3) B-spline trajectory, per-frame exposure times, and a learned CRF in one differentiable pipeline. The synthetic evaluations are strong, and the ablations show each component earns its place. The real dataset is a useful addition, and the paper is honest about the scale ambiguity in exposure times.\n\nThe main soft spot is the one the stress-test flags: exposure-time identifiability. The per-image white balance gains and the per-channel tone-mapping MLPs can absorb frame-to-frame brightness differences, and the exposure-normalized loss explicitly removes the absolute brightness constraint. The paper's only quantitative evidence for exposure recovery is rank correlations on two real sequences, which say nothing about absolute accuracy. Since the synthetic Blender scenes have ground-truth exposure times, the authors should have reported absolute errors there. That omission is noticeable but not disqualifying: the exposure time also controls the blur integral, and the synthetic deblurring gains with random exposure initialization suggest the model is not just fitting brightness. Still, without absolute error numbers or an identifiability analysis, the central claim is only partially supported.\n\nTwo smaller issues: real-world novel-view evaluation uses only manually selected sharp frames, so it doesn't directly test the blur-and-exposure setting in the wild; and there are no error bars anywhere. The paper also doesn't ship code or data, so independent verification is not possible right now. I would not call any of this fatal. The design is sound, the writing is clear, and the contribution is a genuine synthesis rather than a repackaging.\n\nFor a reader: this is for people working on HDR 3D reconstruction, deblurring, or robust NVS from unconstrained video. It deserves a serious referee. My recommendation would be to accept with major revision: require synthetic exposure-time absolute error, an ablation that removes per-image WB or constrains it, and ideally release code/data.","headline":"Casual3DHDR is a real step forward for HDR 3D reconstruction from casual video, but the exposure-time identifiability evidence is thinner than the central claim warrants.","tokens_in":17917,"tokens_out":2896,"would_cite":true,"duration_ms":28559,"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":"Casual3DHDR claims that a handheld auto-exposure video, with unknown exposure times and motion blur, is sufficient input to reconstruct a high-dynamic-range 3D scene by jointly optimizing exposure, camera trajectory, and the camera…","keywords":["high dynamic range","3D Gaussian splatting","motion deblurring","casual video","exposure time estimation","camera response function","continuous-time trajectory","novel view synthesis"],"falsifier":"Run the method on a synthetic sequence with known ground-truth exposures, camera response curve, and HDR radiance, then compare the recovered HDR radiance and exposure times against ground truth in absolute units, not just rank correlation. A decisive test is to construct two different ground-truth camera-response/exposure combinations that produce nearly identical LDR frames and check whether optimization converges to the true combination or to an appearance-equivalent one.","tokens_in":16895,"feed_emoji":"🎥","tokens_out":6896,"duration_ms":66355,"temperature":0.7,"pith_summary":"The paper claims that a casually recorded auto-exposure video—handheld, blurred, and with unknown, changing shutter speeds—is enough input to reconstruct a high-dynamic-range (HDR) 3D scene. It proposes a single pipeline that treats each video frame as the output of a physical imaging process: scene radiance viewed along a continuous camera trajectory, accumulated over an unknown exposure time, then mapped through a learned camera response function. Exposure time, trajectory, response curve, and the 3D Gaussian scene are optimized together, so no calibrated exposure readings or static tripod shots are needed. On synthetic and real casual videos, the method reports higher novel-view quality, deblurring quality, and pose accuracy than prior HDR and deblurring reconstruction approaches.","feed_headline":"Blurry auto-exposure video is enough to build 3D HDR scenes","feed_subtitle":"Motion blur and brightness come from one exposure time, so no shutter-speed metadata or tripod is needed.","key_machinery":"The load-bearing object is the unified physical image-formation model $\\mathbf{B}(x,y) = F\\left(\\int \\mathbf{H}(x,y,P(t))\\, dt\\right)$, with $P(t)$ supplied by a cumulative SE(3) B-spline, a smooth curve through control knots that yields a camera pose at any timestamp. The spline makes camera motion during exposure differentiable, so blur is rendered as the average of virtual sharp views. Exposure time is the coupling variable: it determines how many virtual views the integral averages over and scales the accumulated irradiance before the camera response function maps it to pixel values. A per-channel MLP tone mapper plus an image-varying white-balance vector plays the role of $F$. Jointly optimizing all of these makes the blur width and brightness level mutually consistent, which is how the method separates scene radiance from camera state.","core_discovery":"The central claim is that motion blur and exposure variation are not separate nuisances but two readings of the same physical quantity: the exposure interval. Along a continuous-time SE(3) B-spline trajectory, each observed low-dynamic-range frame is modeled as $\\mathbf{B} = F\\left(\\int \\mathbf{H}(x, y, P(t))\\, dt\\right)$, where $\\mathbf{H}$ is the HDR radiance field, the integral averages over camera motion during the shutter interval to produce blur, and $F$ is a camera response function with white balance and per-channel tone mapping to produce brightness. Because exposure time controls both the width of the blur kernel and the total accumulated light, optimizing it couples deblurring with HDR reconstruction. The method learns the 3D HDR scene, the trajectory, the per-frame exposure times, and the camera response function jointly, and reports that this recovers usable exposure times, sharp frames, re-exposable HDR renders, and accurate poses even when exposure times are initialized randomly.","pith_inferences":["Editorial inference: If exposure recovery is only up to a global scale, as the appendix's correlation-based evaluation suggests, then the recovered HDR radiance is also known only up to an unknown multiplier; relative re-exposure editing survives, but absolute radiometric comparison across scenes does not.","Editorial inference: Adding an explicit prior that keeps the learned tone-mapping curve smooth and monotone, or including a single calibration frame with known shutter speed, would likely convert the current correlation-level exposure recovery into absolute exposure calibration; that is a testable next step.","Editorial inference: The same coupling of blur width and brightness through one timing parameter transfers naturally to rolling-shutter readout, event-camera capture, or mixed flash/no-flash video, where the timing parameter plays an analogous role."],"forward_implications":["A consumer camera's video mode, with auto-exposure on, becomes a sufficient input for 3D HDR reconstruction, with no exposure metadata, no RAW capture, and no tripod required.","The same reconstruction yields sharp images at any chosen exposure, so deblurring and HDR exposure editing are derived outputs of one optimization rather than separate tasks.","Because poses, exposure times, and the camera response curve are optimized together instead of in a separate preprocessing step, scenes where feature-based pose estimation fails under brightness changes and blur can still be reconstructed.","Randomly initialized exposure times converge to values that track the true exposure times, so the pipeline can run without a calibration step.","The continuous trajectory representation also provides camera poses at arbitrary timestamps, enabling high frame-rate novel-view rendering from low frame-rate input video."],"supporting_citations":[{"why":"Defines the 3D Gaussian scene representation and the differentiable rasterization primitive that the method optimizes.","marker":"[18]"},{"why":"Provides the cumulative SE(3) B-spline interpolation formula that yields camera poses at arbitrary timestamps during exposure.","marker":"[32]"},{"why":"Supplies the physical motion-blur model that averages sharp views along an estimated trajectory, the backbone this method extends to HDR and exposure optimization.","marker":"[53]"},{"why":"Establishes the HDR radiance-field task with camera response function estimation from multi-exposure LDR inputs, the setting this paper relaxes by removing known exposures.","marker":"[14]"},{"why":"Self-calibrating HDR radiance-field baseline that learns the camera response function but assumes sharp images and known exposures; used as a comparison point.","marker":"[17]"},{"why":"Models per-image appearance variation in 3D Gaussian splatting and serves as a comparison method for brightness-varying inputs.","marker":"[61]"},{"why":"Deblurring 3D Gaussian splatting baseline via bundle-adjusted trajectory estimation; comparison for deblurring, pose, and novel-view quality.","marker":"[63]"},{"why":"Provides the rasterizer and optimization base used to implement the method.","marker":"[58]"}],"fun_headline_variants":["Blur and exposure share one dial for 3D HDR","Casual video to re-exposable 3D HDR, one integral","Deblur and tone-map in one splatting pass","Exposure time couples deblurring and HDR in 3DGS","One integral: blur and brightness from exposure time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that, given only low-dynamic-range frames, the exposure times and camera response curve can be recovered as physical quantities even though the tone-mapping network and white-balance terms are also free to change; if those flexible components absorb the brightness differences, the optimized exposure times and HDR radiance might fit the images without being the true physical values.","fun_headline_variants_meta":{"raw":{"variants":["Blur and exposure share one dial for 3D HDR","Casual video to re-exposable 3D HDR, one integral","Deblur and tone-map in one splatting pass","Exposure time couples deblurring and HDR in 3DGS","One integral: blur and brightness from exposure time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000772,"raw_usage":{"total_tokens":3441,"prompt_tokens":993,"completion_tokens":2448,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":2357}},"tokens_in":609,"tokens_out":2448,"duration_ms":16387,"temperature":1.0,"reasoning_tokens":2357,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:32:52.788088+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on a synthetic sequence with known ground-truth exposures, camera response curve, and HDR radiance, then compare the recovered HDR radiance and exposure times against ground truth in absolute units, not just rank correlation. A decisive test is to construct two different ground-truth camera-response/exposure combinations that produce nearly identical LDR frames and check whether optimization converges to the true combination or to an appearance-equivalent one.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the cumulative SE(3) B-spline interpolation formula that yields camera poses at arbitrary timestamps during exposure."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the HDR radiance-field task with camera response function estimation from multi-exposure LDR inputs, the setting this paper relaxes by removing known exposures."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Self-calibrating HDR radiance-field baseline that learns the camera response function but assumes sharp images and known exposures; used as a comparison point."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Deblurring 3D Gaussian splatting baseline via bundle-adjusted trajectory estimation; comparison for deblurring, pose, and novel-view quality."}],"review_version":1}