{"id":"bd73fc16-2d82-4eea-9f3a-7098feb99f4e","arxiv_id":"2412.02950","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Ceiling-DSO applies the existing DSO algorithm to an upward-facing camera, and the authors tune its parameters to obtain what they call acceptable error rates on two indoor industrial sequences.","lead":"This paper mounts an upward-facing camera on an indoor mobile robot and runs Direct Sparse Odometry (DSO) on ceiling images to estimate the robot's motion. It tunes DSO parameters and reports that 15 frames per second with a window size of 7 gives an acceptable balance of accuracy and real-time speed on two test sequences.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (12)'s linear-response/no-vignetting assumption is unvalidated; DSO's affine brightness correction cannot compensate spatially varying vignetting, so the reported error rates may be specific to this uncalibrated setup and need a calibration ablation.","rationale":"The paper's central claim is that DSO's photometric direct tracking transfers to ceiling imagery without assumptions about ceiling structure. That claim is plausible: the experiments show ceiling trajectories that roughly match LiDAR-SLAM ground truth, and DSO is an established algorithm with per-frame affine brightness parameters. The key scientific risk is not the absence of an explicit ceiling model but the unverified photometric simplification in Eq. (12). The RealSense cameras used in the experiments are known to have sRGB-like nonlinear response and vignetting; DSO's residual model (Eq. 4) only absorbs per-frame affine changes, so a spatially varying violation biases depth estimation and point selection. Since neither code nor dataset is released, this assumption cannot be checked by readers. I agree with the reader's identification of this as the weakest assumption. The concern does not refute the claim; it narrows it: Ceiling-DSO as evaluated is tied to an uncalibrated camera model. A calibration ablation would settle whether the reported 'acceptable error rates' change materially. If they do, the claim should be conditioned on photometric calibration; if they do not, the simplification is benign and the conditional acceptance can be upgraded. Thus the reader's CONDITIONAL verdict should remain unchanged pending that check.","tokens_in":12642,"tokens_out":8447,"duration_ms":94604,"concrete_test":"Run the two reported sequences through DSO twice: once with raw frames as in the paper (Eq. 12) and once after applying photometric calibration (estimate G and V with a radiometric calibration target or DSO's photometric calibration tool), keeping all other parameters identical to the reported best settings. Compare the relative position error (and RPE) as a function of traveled distance. If calibration changes the median error by more than about 20% or systematically reduces residuals near image edges, the Eq. (12) simplification is a first-order effect and the current 'acceptable error rates' are not robust to photometric unmodeling.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 (Eq. 12) replaces the DSO photometric model with G(x)=x and V(x)=1, and Section 3 uses raw images from Intel RealSense D455/D435i cameras. RealSense RGB sensors have a nonlinear (approximately sRGB) response and measurable vignetting; these are spatially varying and cannot be captured by the per-frame affine brightness parameters (a_i,b_i) in Eq. (4). If the mismatch is non-negligible, photometric residuals are biased by a function of pixel radius and gradient direction, and DSO's point selection and inverse-depth estimates can absorb that bias, inflating or concealing trajectory error. The central claim is 'acceptable error rates,' but no evidence is given that the reported errors are not an artifact of this unmodeled photometric transfer function. This is load-bearing because Eq. (12) is the only photometric calibration step in the Ceiling-DSO implementation, and the paper does not release code, data, or a calibration ablation to test it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents Ceiling-DSO, an adaptation of Direct Sparse Odometry (DSO) to an upward-facing monocular camera for indoor industrial robot localization. It claims that DSO's photometric tracking works on ceilings without assumptions about landmarks or shapes, and reports 'acceptable error rates' compared to LiDAR-SLAM ground truth. The authors created a custom dataset with an industrial robot, systematically varied image size, frame rate, and optimization window size, and provide qualitative trajectory plots, relative-error plots, box plots, and runtime tables for two sequences.","tokens_in":12890,"tokens_out":4251,"duration_ms":40765,"significance":"If the claims hold, Ceiling-DSO would be a useful, assumption-free ceiling-vision odometry candidate for dynamic industrial environments, avoiding moving-object interference. The paper builds on a mature algorithm (DSO), so the core mechanism is plausible; its main contribution is an empirical demonstration and parameter analysis. Strengths include the real-world dataset, systematic evaluation across 24 parameter combinations per sequence, and explicit runtime and speed-factor analysis. However, the evidence base is narrow (two sequences, one building, no error bars), the photometric model is simplified without validation, no code or data are released, and the evaluation methodology leaves several load-bearing questions open. The contribution is incremental relative to the authors' ICARCV 2022 paper [1].","major_comments":[{"comment":"The implementation assumes a linear camera response and no vignetting (G(x)=x, V(x)=1), yet the experiments use Intel RealSense D435i/D455 cameras, whose nonlinear response and lens vignetting are not calibrated or compensated. The affine brightness parameters (a_i,b_i) in Eq. (4) cannot model spatially varying vignetting, so the photometric residuals are systematically biased if the assumption is false. Since the central claim is 'acceptable error rates', the paper must either provide photometric calibration of the cameras or include an ablation demonstrating that the reported errors are insensitive to this mismatch.","section":"Section 2, Eq. (12)"},{"comment":"The quantitative evaluation relies entirely on qualitative trajectory plots, relative-error plots, and box plots; no overall numerical metrics (e.g., RMSE, mean/median relative error, maximum error) are reported in the text or tables. As a result, the abstract's claim of 'acceptable error rates' is not quantitatively defined or reproducible. Please report standard trajectory-error metrics for each parameter combination and sequence.","section":"Section 4, Figs. 8-10"},{"comment":"The ground truth trajectory is produced by LaMa SLAM [30] from the safety LiDAR, but the accuracy of this SLAM solution is not quantified, and no loop-closure or map-quality assessment is provided. If LaMa's own drift is comparable to the reported relative errors, the comparison is inconclusive. The authors should report the LiDAR-SLAM trajectory accuracy (e.g., against known ground-truth geometry or a second SLAM method) or at least discuss its expected drift.","section":"Section 3.3"},{"comment":"The recommended parameters (15 fps, window size 7) are selected by inspecting the same two sequences used for the reported performance, without a held-out validation set; this introduces a selection bias and overstates generalizability. Additionally, the monocular scale lambda is fitted during trajectory alignment in Eq. (14), so trajectory scale error is not evaluated; the paper should report how sensitive the error rates are to the scale estimate and validate parameters on unseen sequences.","section":"Section 4"}],"minor_comments":[{"comment":"There are numerous language errors (e.g., 'an ceiling-vision', 'On one hand', 'This used this embedded computer') that should be corrected.","section":"Throughout"},{"comment":"The text first states that direct methods use information from all pixels, then later describes DSO as sparse; this taxonomy should be clarified to avoid confusion.","section":"Section 2"},{"comment":"The synchronization definitions of G' and P' are notationally confusing between sets and ordered sequences; please make the matching explicit.","section":"Eq. (13)"},{"comment":"The text refers to 'plot 10' instead of 'Figure 10', and several trajectory figures lack legends within some panels; please standardize figure captions and legends.","section":"Section 4"},{"comment":"The column heading 'Win.' should be expanded to 'Window size' for clarity, and the tables would benefit from a note that timings are averages over the tested sequences.","section":"Tables 1-2"}],"recommendation":"major_revision","confidential_remarks":"This is an engineering application paper whose novelty over the authors' ICARCV 2022 paper [1] appears limited to a more detailed parameter study and analysis. The lack of code and data release, combined with the narrow two-sequence evaluation, makes it difficult to assess reproducibility. Given the journal's scope, the editors may wish to consider whether the incremental contribution and the current level of evidence justify publication after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper extends the authors' own ICARCV 2022 work, applying Direct Sparse Odometry to an upward-facing ceiling camera. What's new is a systematic parameter sweep on two sequences, looking at image resolution, frame rate, and optimization window size, with runtime and accuracy trade-offs. That is a legitimate, if incremental, contribution for practitioners who want a quick ceiling-based odometry option without landmark assumptions.\n\nThe writing is clear, the DSO background is competently summarized, and they are explicit about the simplifications they make. The runtime analysis is useful, and the qualitative trajectory plots suggest the approach basically works. I give them credit for being upfront that this is an extension and for discussing future work like scale estimation and loop closing.\n\nThe soft spots are real but not necessarily fatal. The evaluation is thin: two sequences from one building, no repeated runs, no error bars, no numerical error summary in the text, and no comparison against a baseline like feature-based ceiling SLAM or DSO with a forward camera. The photometric calibration assumption, Eq. (12), is a genuine concern. RealSense cameras have a nonlinear response and vignetting, and DSO's per-frame affine brightness parameters cannot absorb spatially varying vignetting. The stress-test note is right that this could bias photometric residuals. But DSO's robust Huber norm and gradient-based point selection might make the system forgiving enough, so I would not call it a fatal flaw. Still, the authors should run a calibration ablation or at least discuss why the simplification is acceptable. The parameter tuning on the evaluation sequences is a minor circularity, typical for this kind of study, but they should acknowledge it. The LiDAR ground truth has its own unquantified drift, which is worth a sentence.\n\nWho is this for? Researchers and engineers working on indoor mobile robot localization, especially in dynamic environments where looking at the ceiling is attractive. It is a plausible engineering result but not a breakthrough. It deserves peer review because it fills a practical niche and the parameter study has some value, but the referees should push for stronger evidence.\n\nRecommendation: send it to review, but expect major revisions on evaluation robustness, baseline comparisons, and the photometric calibration question.","headline":"A credible but incremental ceiling-DSO parameter study; the evaluation is thin and the unvalidated photometric calibration assumption is a real soft spot, but the core idea is sound enough to referee.","tokens_in":13395,"tokens_out":1699,"would_cite":false,"duration_ms":18977,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T40","68T45","68U10"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that Direct Sparse Odometry, run on an upward-facing camera, can track a mobile robot's trajectory from the ceiling alone, with no assumptions about ceiling shapes or landmarks, and that the resulting relative errors are…","keywords":["Ceiling vision","Direct Sparse Odometry","Visual odometry","Mobile robot localization","Indoor industrial environments","Upward-facing camera","Photometric model"],"falsifier":"Record a ceiling sequence with a camera known to have strong vignetting or a highly nonlinear response curve, run Ceiling-DSO exactly as described (without photometric calibration), and check whether the relative trajectory error stays in the low-centimeter range reported in the paper; if the error grows substantially or tracking fails, the claim that the simplified photometric model suffices for arbitrary ceilings is falsified. A second check is a nearly textureless white ceiling, which would test whether DSO's sparse high-gradient point selection really needs no landmarks at all.","tokens_in":12470,"feed_emoji":"🤖","tokens_out":3917,"duration_ms":39185,"temperature":0.7,"pith_summary":"Ceiling-DSO repurposes Direct Sparse Odometry (DSO) for ceiling vision: a camera pointed upward feeds raw images into DSO's photometric optimization, so the robot tracks its motion from ceiling texture without detecting lamps, corners, doors, or artificial markers. The payoff for indoor industrial robots is localization that ignores crowded, dynamic floor scenes and works on whatever the ceiling happens to look like. The paper validates this with a custom real-world dataset and a LiDAR-based ground truth, and it identifies parameter settings that keep the system real-time. If the claim holds, an ordinary upward camera plus a direct odometry algorithm becomes a generic, low-infrastructure indoor localization solution.","feed_headline":"Ceiling-only camera tracks indoor robots with no markers","feed_subtitle":"Direct photometric tracking of the ceiling keeps indoor robot odometry accurate with no landmark assumptions.","key_machinery":"The load-bearing mechanism is DSO's photometric error, the weighted sum of Huber-norm residuals comparing raw pixel intensities of a sparse set of high-gradient points across keyframes, minimized by Gauss-Newton on the Lie algebra se(3) within a bounded optimization window. The paper deliberately simplifies the photometric image formation model by assuming a linear camera response and no vignetting (G(x)=x and V(x)=1 everywhere), so the brightness transfer is modeled only as an affine exposure correction. This direct, intensity-based optimization is what removes the need to detect or classify ceiling features, and it is also the component whose accuracy depends on the camera obeying the simplified photometric model.","core_discovery":"On the paper's own terms, the central discovery is that DSO's direct photometric formulation transfers to ceiling imagery essentially as-is: minimizing the weighted Huber-norm photometric residuals over a sliding window of keyframes is sufficient to estimate ego-motion from a ceiling, even though the scene is far above the camera and contains no engineered landmarks. The paper shows this empirically on two real trajectories in a 21 m by 15 m space with an inclined ceiling between 4 and 6 meters high, reporting relative errors that stay in the low-centimeter range for most parameter combinations and remain within acceptable bounds even below the optimal settings. It also reports that frame rate is the dominant accuracy-relevant parameter, that downscaling the image resolution does not hurt accuracy, and that a maximum optimization window of 7 keyframes offers the best accuracy-to-runtime tradeoff.","pith_inferences":["Because the system already works with a deliberately simplified photometric model (linear response, no vignetting), applying true photometric calibration to the RealSense cameras would likely reduce the reported errors further; the paper does not test this direction.","The finding that low frame rates hurt accuracy mainly during curved motion suggests an adaptive frame-rate policy (low on straightaways, high in turns) as a natural extension, building on the paper's own observation.","Once the authors release their custom dataset, it could become a shared benchmark for ceiling-vision odometry, filling the gap the paper identifies of having no public dataset for this task.","Ceiling scenes are static geometrically but not photometrically (windows, lamps, shadows), so a direct test of robustness to changing illumination would be a quick way to probe how far the no-assumption claim extends."],"forward_implications":["Indoor mobile robots can localize by looking up, avoiding the need to filter out moving people, robots, and forklifts from the camera image.","Deploying ceiling-vision odometry in new buildings does not require placing markers, measuring ceiling landmarks, or assuming a particular ceiling type.","The parameter study gives concrete operating guidance: run at about 15 frames per second with a window of 7 keyframes, and consider reducing image resolution to 424 x 240 to save compute without losing accuracy.","The real-time speed factors reported in the paper indicate that the system can run online on an embedded industrial PC for most parameter choices, with the 30 fps / window-15 combination being the main exception.","Future fusion with wheel encoders could supply the metric scale that monocular DSO lacks, turning Ceiling-DSO into a complete, drift-bound localization system."],"supporting_citations":[{"why":"Supplies the DSO formulation, photometric error, and Gauss-Newton optimization that Ceiling-DSO adapts to ceiling imagery.","marker":"[12]"},{"why":"Provides the photometric image formation model (response function G, vignetting V) that the paper inherits and then simplifies to linear/no-vignetting.","marker":"[26]"},{"why":"LaMa SLAM is run on the robot's LiDAR data to produce the ground-truth trajectories used for evaluation.","marker":"[30]"},{"why":"Gives the trajectory synchronization, similarity-transformation alignment, and relative-error metrics used to compare Ceiling-DSO against ground truth.","marker":"[31]"},{"why":"The earlier ICARCV version of this system whose evaluation this paper extends with a broad parameter study.","marker":"[1]"},{"why":"The first ceiling-vision SLAM approach, which the paper contrasts with by pointing out that Ceiling-DSO avoids landmark assumptions.","marker":"[18]"}],"fun_headline_variants":["Upward camera turns ceiling into robot's GPS","No markers, no landmarks: ceiling-only odometry","DSO on the ceiling: direct photometric robot tracking","Ceiling-ward camera keeps indoor robots on track","Watch the ceiling: robust indoor robot tracking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The implementation assumes the camera's brightness response is linear and the lens has no vignetting, but the Intel RealSense cameras used have a nonlinear response and vignetting, so the photometric model that DSO optimizes is systematically wrong and the reported error rates may not transfer to other cameras or ceilings.","fun_headline_variants_meta":{"raw":{"variants":["Upward camera turns ceiling into robot's GPS","No markers, no landmarks: ceiling-only odometry","DSO on the ceiling: direct photometric robot tracking","Ceiling-ward camera keeps indoor robots on track","Watch the ceiling: robust indoor robot tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1416,"prompt_tokens":900,"completion_tokens":516,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":442}},"tokens_in":516,"tokens_out":516,"duration_ms":5667,"temperature":1.0,"reasoning_tokens":442,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:55:15.612914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record a ceiling sequence with a camera known to have strong vignetting or a highly nonlinear response curve, run Ceiling-DSO exactly as described (without photometric calibration), and check whether the relative trajectory error stays in the low-centimeter range reported in the paper; if the error grows substantially or tracking fails, the claim that the simplified photometric model suffices for arbitrary ceilings is falsified. A second check is a nearly textureless white ceiling, which would test whether DSO's sparse high-gradient point selection really needs no landmarks at all.","supporting_citations":[{"cited_title":"Evaluation of a Novel DSO-based Indoor Ceiling-Vision Odometry System","cited_arxiv_id":null,"evidence_quote":"The earlier ICARCV version of this system whose evaluation this paper extends with a broad parameter study."}],"review_version":1}