{"id":"75f1828b-21fa-46cf-ab21-01750e0334a1","arxiv_id":"1908.07732","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A processed antique stereo dataset with depth maps and a real-time mobile view synthesis method built on double reprojection inpainting.","lead":"KeystoneDepth converts thousands of antique stereo photographs into rectified pairs with depth, letting viewers move their heads to see parallax in historical scenes. A companion view synthesis technique renders this in real time on mobile devices, turning century-old images into explorable 3D windows.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The dataset's calibration rests on a single hand-set 45° vertical field of view; because uncalibrated rectification leaves the projective frame underdetermined, a wrong FOV introduces non-uniform depth distortion that propagates into every depth map and synthesized view.","rationale":"The reader's weakest-assumption analysis pinpoints the 45° vertical FOV as the most fragile link, and my stress test agrees. The paper's central promises are a calibrated, depth-annotated dataset and a method that extrapolates views 'significantly beyond the input range.' Both depend on the projective reconstruction induced by the assumed intrinsics. If the FOV is wrong, the depth maps are not merely scaled; they are projectively distorted, so the claimed physical realism of the view synthesis is unsupported. The paper acknowledges the assumption but offers only informal plausibility, no sensitivity analysis, and no independent validation. This is a correctness risk, not a disagreement with consensus: it is an internal gap between the strength of the dataset claim and the evidence for the chosen constant. The conditional verdict remains appropriate because the concern could be resolved by a sensitivity analysis or independent FOV estimation; the dataset could still be valuable as a visualization resource even if the calibration is imperfect. I did not elevate the verdict to REJECT because the synthesis method's qualitative performance is demonstrated and the FOV error may be modest in practice. The concrete test would settle whether the assumption actually changes outputs in a material way.","tokens_in":10797,"tokens_out":7918,"duration_ms":86193,"concrete_test":"Render a small set of synthetic antique-style stereo pairs from known 3D scenes with vertical FOVs of 30°, 40°, 50°, and 60°, with realistic texture and noise; run the exact KeystoneDepth rectification (Loop–Zhang plus the 45° FOV assumption) and depth pipeline on each; compare the recovered depth maps and the Section 4.1 quad-corner synthesized views to ground truth using mean log depth error and a perceptual metric such as LPIPS on the rendered views. If errors vary substantially across FOVs or if synthesized-view quality degrades as the true FOV moves away from 45°, the assumption is load-bearing and the dataset's calibration claims need re-scoping; if results are nearly invariant to FOV, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is the hand-set vertical field of view in Section 3.2. The paper fixes the camera intrinsic matrix by assuming a 45° vertical FOV, then computes all disparity maps via FlowNet2 and all quad-corner viewpoints in Section 4.1 from this calibration. Because the rectification is uncalibrated (Loop–Zhang), the pair (focal length, baseline) is not separately observable from the images; it is determined only up to a projective transformation. Choosing a different FOV therefore does not merely rescale all depths uniformly — it changes the ratio of lateral to depth coordinates, so a scene reconstructed under the wrong FOV is a non-metric (projective) distortion of the true scene. As a consequence, every depth map in the released dataset inherits this arbitrary choice, and the 'extrapolate significantly beyond the input range' results in Section 6 are generated from a geometry that is not physically grounded. The paper sets the parameter by informal plausibility ('most plausible depth scaling in our experiments') and provides no sensitivity analysis, no comparison to independently estimated FOVs (e.g., from vanishing points or documented Keystone camera models), and no error bars. Since the primary contribution is the calibrated dataset itself, this single unsupported constant is the weakest link in the chain from raw scans to claimed 3D views.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript describes the KeystoneDepth dataset and a view-synthesis system for antique stereographs. The authors downloaded 29,480 stereographs from the Keystone-Mast collection, automatically filtered and crowd-cropped them, rectified each surviving pair with Loop–Zhang rectification, estimated disparities with FlowNet2, and retained 10,134 processed pairs. Each dataset entry is intended to include metadata, bounding boxes, a rectified stereo pair, camera parameters, and disparity/depth maps. The second contribution is a representation built from one reference intensity-plus-depth image and four synthesized quad-corner views, with holes filled by a partial-convolution inpainting network trained using a new double-reprojection self-supervision scheme. A mobile AR application, KeystoneAR, is presented. The paper includes a manual evaluation of 711 samples (23% very few artifacts, 49% some artifacts, 28% failures) and a user study comparing the approach with Stereo Magnification.","tokens_in":11127,"tokens_out":8317,"duration_ms":88044,"significance":"If the calibration and depth maps are trustworthy, the dataset would be a substantial and unusual resource for historical stereo, rectification, and view-synthesis research. The paper is commendably detailed about its processing pipeline, openly reports its failure rates rather than cherry-picking successes, and includes a user study against a published baseline. The main correctness risk is the unvalidated calibration assumption in Section 3.2, which propagates into every depth map and synthesized view; in addition, the abstract overstates the processed dataset size, and the real-time claim is not backed by measurements. These issues are fixable with additional validation and by aligning the claims with the presented evidence, so the work is worthy of major revision rather than rejection.","major_comments":[{"comment":"The vertical field of view is set by hand to 45 degrees because it produced \"the most plausible depth scaling in our experiments.\" Since the rectification is uncalibrated and the disparity maps are obtained from optical flow, the camera intrinsics and baseline are not separately observable from the images; this assumed FOV therefore enters every calibration record, every depth map, and every view-extrapolation range in Section 4.1 and Section 6. Please provide a sensitivity analysis showing how depth maps and synthesized views change as the FOV is varied over a plausible range, or obtain an independent FOV estimate (e.g., from vanishing points or documented Keystone camera models). At a minimum, the dataset documentation should state explicitly that the depths are not metric and that the FOV is an unvalidated assumption rather than a measured quantity.","section":"3.2"},{"comment":"The abstract and introduction describe the collection as consisting of \"tens of thousands of stereographs,\" but Section 3 reports that 29,480 stereographs were considered and disparity maps were computed for only 10,134 pairs. Moreover, the manual evaluation in Section 6 finds that 28% of 711 sampled entries are failures and only 23% have very few artifacts. The current wording overstates the actual processed resource and should be aligned with the 10,134 number, or the text should clearly distinguish raw scans from processed entries and present the projected expansion as a plan.","section":"Abstract and Section 3"},{"comment":"The abstract and Section 4 claim that the view-synthesis approach runs \"at real-time rates on a mobile device,\" and Section 7 describes the KeystoneAR app, but Section 6 contains no timing measurements. Please report measured frame rates on a specific device, including the resolution and the number of GD images rendered, and if possible compare timings with the Stereo Magnification baseline. Without such measurements the real-time claim is unsupported.","section":"4 and 6"},{"comment":"The notation around the hole mask is inconsistent. The text defines M as \"a binary mask corresponding to the holes that are to be inpainted,\" yet the network inputs are written as (I⊙M, D⊙M, B⊙M), and the losses are defined as L_valid = ||(Dp−Dt)⊙M||1 and L_hole = ||(Dp−Dt)⊙(1−M)||1. If M=1 marks holes, then the input multiplication zeroes out the valid region and L_valid is actually computed in the hole region; if M=1 marks valid pixels, the textual definition is wrong. Please make the mask convention explicit and consistent throughout, since this is essential for reproducing the training procedure.","section":"5.3"}],"minor_comments":[{"comment":"The introduction and Figure 1 caption say \"over 10,000 rectified antique stereographs,\" while the abstract says \"tens of thousands\"; unify the terminology so that the reader can tell when raw scans versus processed entries are being described.","section":"1 and Figure 1"},{"comment":"The formula rw = rh = 96b/dmax × √2/2 uses b, the camera baseline, but the paper does not explain how b is obtained from uncalibrated rectification and optical flow. Please define all variables and state whether b is in pixels, metric units, or arbitrary units.","section":"4.1"},{"comment":"There are several typographical errors: \"a batch size if4\" should be \"a batch size of 4,\" \"novel novel lightweight\" in the conclusion repeats \"novel,\" and the two references numbered [20] and [21] are the same paper and should be merged.","section":"5.4 and 8"},{"comment":"The abbreviation \"GD image\" is used in the caption of Figure 3 before it is defined in the text; define the term at its first occurrence in Section 4.1 or in the caption itself.","section":"4.1"},{"comment":"The user study reports only mean preference percentages (62.3% vs 27.8%). Please report the number of comparisons, the per-scene variance, and a confidence interval or statistical test so the reader can assess the strength of the preference.","section":"6"}],"recommendation":"major_revision","confidential_remarks":"The dataset is the most valuable contribution, but the unvalidated 45-degree FOV assumption in Section 3.2 is the kind of issue that can undermine the advertised calibration and depth products if not addressed with sensitivity analysis or explicit non-metric caveats. The runtime claim also needs a quantitative basis. The novelty of the view-synthesis component is moderate, but the dataset itself is likely to have lasting impact."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version. This paper is worth your time if you care about stereo datasets or historical imaging: it assembles a large collection of rectified antique stereographs with depth maps, which is genuinely new. The view-synthesis pipeline is decent, but the calibration sits on a hand-set 45° vertical FOV, and that assumption propagates into every depth map and synthesized view. The abstract also overstates the collection size.\n\nWhat's actually new: the KeystoneDepth dataset — 10,134 rectified pairs with disparity maps, metadata, and an honest breakdown of failure cases (28% in a manual evaluation). No one has released anything like this. The double-reprojection method for training an inpainting network on realistic disocclusion holes is a clever low-cost idea, and the boundary mask helps preserve silhouette edges. The mobile AR demo is a nice practical result.\n\nSoft spots, in order of importance. First, the FOV issue. The stress-test note is correct: with uncalibrated rectification, the focal length and baseline are not separately identifiable. Setting the vertical FOV to 45° doesn't just scale depths uniformly; it changes the ratio of depth to lateral dimensions. If the true FOV is different, every depth map is a non-metric (affine) distortion of the scene. The paper says this produces 'the most plausible depth scaling,' which is an informal heuristic, and it provides no sensitivity analysis or independent FOV check. For a dataset meant as a research resource, that's a real limitation; for the visualization demo, it is tuned to look good and is comparatively benign.\n\nSecond, the abstract says 'tens of thousands' but the actual processed count is 10,134, which is still large but not tens of thousands. Third, the real-time claim has no timing numbers — no FPS or latency. Fourth, the user study preference is reported without statistical significance. Fifth, no code or data link appears in the paper, which hampers reproducibility and immediate use.\n\nOverall: the dataset is a real contribution, the method is reasonable, and the paper is more candid about failures than most. The FOV caveat is important and should be disclosed prominently and ideally corrected with a robust FOV estimator. For metric uses of the dataset, that's a concern; for the stated cultural-heritage visualization goal, it's not fatal. I'd send this to peer review — the dataset alone justifies it — and I'd ask for data release, timing numbers, and a sensitivity analysis on the FOV.","headline":"A genuinely useful historical stereo dataset, but the hand-set 45° FOV puts an affine distortion into every depth map and the abstract oversells the size.","tokens_in":11597,"tokens_out":5943,"would_cite":true,"duration_ms":59758,"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":"The paper converts antique stereo photographs into a 10,000-pair calibrated 3D dataset and shows a single stereo pair can drive real-time six-degree-of-freedom view synthesis on a phone.","keywords":["KeystoneDepth dataset","antique stereographs","stereo rectification","depth estimation","view synthesis","image inpainting","augmented reality","optical flow"],"falsifier":"Take stereographs that show objects of known physical size, recover depths with the assumed 45-degree field of view, and check whether the implied proportions and parallax motion are consistent with the known geometry across many scenes; if a different field of view is needed to make depths coherent, the central assumption fails. A complementary test is to render synthetic stereo pairs with known focal lengths and baselines and run the full pipeline, then compare recovered depth ratios and extrapolation ranges to the known values.","tokens_in":10628,"feed_emoji":"🕰️","tokens_out":10171,"duration_ms":90507,"temperature":0.7,"pith_summary":"KeystoneDepth is a new public collection of more than 10,000 rectified (row-aligned) historical stereo photographs, each paired with camera calibration, rectification transforms, and depth maps, recovered from a scanned archive of 250,000 stereographs made between 1860 and 1963. The paper further claims that a single rectified stereo pair is enough to synthesize smooth six-degree-of-freedom views over a head-sized viewing volume at real-time rates on a mobile device. The method represents each scene as a reference grayscale-plus-depth image plus four synthesized corner views, and trains a neural network to inpaint the depth-boundary holes created when viewpoints move. If the claims hold, the field gains its largest and most diverse real stereo dataset for historical scenes, and immersive augmented-reality tours of the past become practical.","feed_headline":"10,000 old stereo photos gain depth and live 3D views","feed_subtitle":"A new public dataset pairs a century of stereographs with depth maps, plus a phone-rendered AR window into history.","key_machinery":"The load-bearing object is the five-image GD scene representation: the left stereo view plus four rendered grayscale-plus-depth corner views, stored as texture-mapped triangle meshes, which together define a quadrilateral viewing volume inside which new viewpoints are rendered in real time. The double-reprojection technique is the training mechanism that makes this representation learnable: a depth-and-intensity mesh is rendered from a corner viewpoint and then back to the reference viewpoint, producing hole masks whose structure matches the disocclusions that real extrapolation creates, so the inpainting network needs no ground-truth novel views. A boundary mask marks foreground-side depth discontinuities and is fed to the network to keep silhouettes sharp, and the network itself is a partial-convolution encoder-decoder trained with separate intensity and depth losses.","core_discovery":"The central claim is that a century of antique stereographs—uncalibrated, damaged, misaligned scans—can be converted into a large calibrated stereo dataset, and that a single rectified stereo pair is sufficient to generate new views over a continuous viewing volume at interactive rates. The pipeline filters the archive by feature matching and manual inspection, rectifies each pair with an uncalibrated rectification method under an assumed 45-degree vertical field of view, and estimates disparity from the horizontal component of dense optical flow. From about 29,480 raw scans it produced depth maps for 10,134 pairs. For rendering, the scene is stored as a reference grayscale-plus-depth image plus four synthesized corner views that define a quadrilateral viewing volume, with the missing regions filled by a partial-convolution network trained through double reprojection. The paper reports that this scheme extrapolates well beyond the input stereo baseline, avoids the depth-quantization artifacts of multiplane-image representations, and was preferred by 62.3% of participants in a blind comparison.","pith_inferences":["An extension the authors do not pursue is rescaling the depth maps using independently known scene dimensions (for example, buildings that still stand) to replace the hand-set field of view with measured calibration; this would turn relative depths into usable metric geometry.","The double-reprojection training scheme is not tied to historical stereo: any single RGB-D image could be double-reprojected to create disocclusion-hole training data, so the method could transfer to modern depth cameras and indoor scenes.","Because the recovered depths are only up to an unknown scale fixed by the assumed field of view, the physical size of the AR 'window' is not known; the experience is convincing in relative parallax but not in absolute metrics.","A stronger validation than the paper's user study would compare views synthesized beyond the quad against real photographs of the same historical sites, but for most scenes no such photographs exist, so metric evaluation will require the known-size test."],"forward_implications":["The released dataset gives stereo and 3D-vision researchers more than 10,000 real historical stereo pairs with calibration, rectification, and depth, a scale and diversity no previous stereo dataset provides.","A single antique stereo pair can drive six-degree-of-freedom view synthesis at real-time rates on a phone, so the same lightweight representation can power AR experiences beyond this collection.","Because double reprojection generates realistic disocclusion holes from a single depth-plus-intensity image, the inpainting network can be trained on the target data itself rather than on ground-truth novel views.","The filtering and rectification pipeline can be rerun as more of the 250,000-item archive is digitized, so the collection can grow beyond its current 10,134 depth-bearing pairs.","The explicit-geometry-plus-inpainting approach avoids the depth-quantization 'stack of cards' artifact of learned multiplane images and supports a larger viewing volume, which is why most blind user-study participants preferred it."],"supporting_citations":[{"why":"supplies the raw stereograph scans and metadata that form the KeystoneDepth dataset.","marker":"[32]"},{"why":"provides the uncalibrated stereo rectification method the pipeline adapts.","marker":"[22]"},{"why":"provides the dense optical-flow network whose horizontal component is used as disparity.","marker":"[13]"},{"why":"supplies the partial-convolution architecture and intensity losses the inpainting networks build on.","marker":"[20]"},{"why":"is the learned multiplane-image view-synthesis method used as the main comparison baseline.","marker":"[37]"},{"why":"motivates the quadrilateral viewing-volume sampling for light-field rendering.","marker":"[17]"},{"why":"supplies the synthetic indoor dataset used to pretrain the inpainting networks.","marker":"[35]"},{"why":"provides the standard encoder-decoder backbone whose convolutions are replaced by partial convolutions.","marker":"[26]"}],"fun_headline_variants":["A century of stereographs gains depth and real-time 3D views","10K antique stereo photos now a calibrated dataset with depth","Mobile AR window into history from 100-year-old stereo pairs","Largest stereo collection: 10K pairs with depth, mobile view synthesis","Real-time 3D views from a single rectified historical stereo pair"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes a vertical field of view of 45 degrees to set each camera's focal length, and that single assumption fixes every depth map and every extrapolated view; if the true field of view differs, the recovered 3D geometry is systematically distorted.","fun_headline_variants_meta":{"raw":{"variants":["A century of stereographs gains depth and real-time 3D views","10K antique stereo photos now a calibrated dataset with depth","Mobile AR window into history from 100-year-old stereo pairs","Largest stereo collection: 10K pairs with depth, mobile view synthesis","Real-time 3D views from a single rectified historical stereo pair"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1318,"prompt_tokens":870,"completion_tokens":448,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":356}},"tokens_in":486,"tokens_out":448,"duration_ms":509906,"temperature":1.0,"reasoning_tokens":356,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:57:05.449061+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take stereographs that show objects of known physical size, recover depths with the assumed 45-degree field of view, and check whether the implied proportions and parallax motion are consistent with the known geometry across many scenes; if a different field of view is needed to make depths coherent, the central assumption fails. A complementary test is to render synthetic stereo pairs with known focal lengths and baselines and run the full pipeline, then compare recovered depth ratios and extrapolation ranges to the known values.","supporting_citations":[{"cited_title":"Keystone- Mast Collection","cited_arxiv_id":null,"evidence_quote":"supplies the raw stereograph scans and metadata that form the KeystoneDepth dataset."},{"cited_title":"Loop and Z","cited_arxiv_id":null,"evidence_quote":"provides the uncalibrated stereo rectification method the pipeline adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the dense optical-flow network whose horizontal component is used as disparity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the learned multiplane-image view-synthesis method used as the main comparison baseline."},{"cited_title":"Levoy and P","cited_arxiv_id":null,"evidence_quote":"motivates the quadrilateral viewing-volume sampling for light-field rendering."},{"cited_title":"Zhang, S","cited_arxiv_id":null,"evidence_quote":"supplies the synthetic indoor dataset used to pretrain the inpainting networks."}],"review_version":1}