{"id":"a0be862e-a6e8-4c72-8461-25fbd8b1f0bb","arxiv_id":"2412.16229","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"TopView predicts a vanishing point with a neural network and builds a homography that maps detected road users into a vectorized bird's eye view without camera calibration.","lead":"This paper presents TopView, a pipeline that learns a scene's vanishing point and uses it to project detected road users from uncalibrated street cameras into a bird's eye view. The authors claim high-precision geolocalisation for urban analytics, but the validation is mostly qualitative and relies on manual calibration.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"A single vanishing point cannot determine a metric homography; the four-point BEV construction plus only global scale/shift leaves anisotropic distortion, so the metric geolocation claim is unsupported.","rationale":"The reader's weakest assumption and my load-bearing concern coincide: the geometric construction in §3.1 assumes that a single learned vanishing point plus two hand-set scalars (z and x) determines a metric orthographic top view. This is not a matter of validation quality alone; it is a fundamental underdetermination. Even with a perfect vanishing-point predictor and perfect detection, the mapping from image to ground plane cannot be made metric without additional information such as camera height, focal length, or known scene distances. The paper's manual calibration tool only provides a global scale and a translation, which are insufficient to correct anisotropic scaling. Therefore the central claims of metric distance measurement and high-precision geolocation from uncalibrated images do not follow from the described method. Other issues the reader identified—the unexplained nuScenes comparison, inconsistent loss description, and missing code—are serious but secondary; fixing them would not repair the geometric gap. I agree with the REJECT verdict, and the proposed synthetic test would quantitatively confirm whether the construction is metric. I credit the paper for a large VP training dataset and a plausible pipeline, but the core geometric claim is not supported.","tokens_in":15605,"tokens_out":4845,"duration_ms":46901,"concrete_test":"Generate synthetic pinhole images of a ground plane with known intrinsics (f, principal point) and extrinsics (height h, pitch θ, roll 0). For two different values of f (or h) that share the same vertical vanishing point, apply the paper's construction: pick two horizontal lines, form the four intersection points, map them to a unit rectangle, and compute the BEV coordinates of known ground-truth points. If the BEV x/y scale ratio changes with f or h, the construction is not metric. Then apply the paper's z/x manual calibration to a second scene; demonstrate that a single global scale and x-shift cannot simultaneously correct distances along both axes. This settles whether the claimed metric BEV is achievable from one VP.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim requires the generated BEV coordinates to be metric up to a global similarity, because the paper advertises distance measurement and GPS-scale geolocation. The Homography module in §3.1 constructs the BEV mapping from a single learned vanishing point plus two hand-chosen horizontal lines, and §4.2 corrects only with a global scale (z) and a global x-shift. For a pinhole camera, a single vertical vanishing point fixes the horizon and camera pitch, but leaves focal length, camera height, and roll undetermined. These unknowns introduce an anisotropic scale between the x and y axes of the ground plane, plus shear/rotation if roll is nonzero. A rectangle of arbitrary aspect ratio in the BEV therefore does not represent equal metric intervals in both directions; a global scalar multiplier cannot fix this anisotropy. The four-point quadrilateral is selected from arbitrary lines through the VP, not from known ground-truth measurements, so the resulting homography is only projective, not metric. Consequently, the claimed high validation in geolocating road users and the social-distancing distances in Application 3 are not entailed by the method as described; the figure overlays appear manually aligned with two degrees of freedom, which cannot compensate for direction-dependent distortion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents TopView, a framework for generating vector bird's-eye-view (BEV) representations of road users from uncalibrated street-level images. It combines a learned vanishing-point estimator (MobileNet backbone), YOLOv5m detection with DeepSORT tracking, a heuristic geometric transformation that converts 2D bounding boxes into 3D bounding boxes using the vanishing point and trajectory lines, and a homography module that maps image points to a BEV vector space using four automatically selected points. The authors demonstrate qualitative applications to traffic analysis, geolocalisation on Google Maps, occupancy analysis, and social-distancing measurement. The central claim is that accurate metric BEV geolocalisation is possible without camera intrinsic or extrinsic parameters.","tokens_in":15936,"tokens_out":6161,"duration_ms":53800,"significance":"If the claimed capability were rigorously established, the work would be of practical interest for urban analytics, CCTV processing, and agent-based modelling, since it avoids camera calibration and outputs vector representations amenable to distance measurement. The paper also bundles several useful engineering components, including detection, tracking, and an anonymised token representation of road users. However, the current manuscript does not substantiate the core geometric claims: no quantitative evaluation of BEV accuracy is provided, the geolocalisation validation is qualitative and partly circular, and the nuScenes results in Table 4 are presented without any protocol. The paper includes no code, data, or trained models, limiting reproducibility. On balance, the significance is real but conditional on substantial additional evidence and corrections.","major_comments":[{"comment":"The homography construction from a single learned vanishing point and two hand-chosen horizontal lines, followed by a manual z/x calibration, does not determine a metric mapping of the ground plane. For a pinhole camera, a single vanishing point fixes the horizon (and pitch) only under additional assumptions about roll, and it leaves focal length, camera height, and roll unconstrained. Consequently, the projective transformation between the image quadrilateral and the BEV rectangle can have an arbitrary anisotropic scale between the two ground-plane axes, and a global scale (z) and x-shift cannot compensate for this direction-dependent distortion. Therefore the paper does not support its advertised capabilities of metric distance measurement (Application 3) or high-precision GPS-scale geolocalisation (Application 1).","section":"§3.1 (Homography) and §4.2 (A simple manual calibration)"},{"comment":"The validation of geolocalisation is circular. Section 4.2 describes a manual calibration tool that adjusts the estimated BEV map by two scalar values (z and x) to align it with Google Maps, and the same Google Maps overlay is then presented (Figs. 1 and 6) as qualitative evidence of 'high precision' localisation. No independent ground-truth coordinates, error metrics in metres, or withheld test sites are provided. Thus the claim of high validation in geolocating road users is not established.","section":"§4.2 and Figs. 1, 6"},{"comment":"The quantitative comparison with nuScenes in Table 4 reports a full set of 3D detection metrics (mATE, mASE, mAOE, mA VE, mAAE, NDS, mAP) for TopView, but the experimental procedure is not described anywhere. It is unclear how the method, which is designed for uncalibrated street-level images, was applied to the nuScenes validation set, which cameras and sequences were used, whether the manual z/x calibration was applied, and what annotations served as ground truth. Without this information, the TopView row is unverifiable and cannot support the claim of competitive 3D detection performance.","section":"Table 4"},{"comment":"Algorithm 1 is under-specified as a method for producing 3D bounding boxes. The algorithm computes an orientation label (e.g., 'turning left' or 'side view') from the intersection of the trajectory line with the top edge of the 2D box, but it does not specify how that label is converted into a concrete 3D bounding box, including its dimensions, the coordinates of its eight corners, or its relationship to the 2D box and the vanishing point. Consequently, the 3D bounding box outputs shown in Fig. 8 are not reproducible from the description.","section":"Algorithm 1"}],"minor_comments":[{"comment":"The architecture figure is referenced as 'Fig. ??' in Section 3.1, but no such figure is present; the reader cannot see the intended data flow.","section":"§3.1"},{"comment":"The text consistently uses 'T able' instead of 'Table' (e.g., 'T able 1', 'T able 4'), and the table captions are inconsistently formatted.","section":"Throughout"},{"comment":"References [20] and [21] are both listed as 'Fast R-CNN'; one should refer to the original R-CNN or the duplicate should be removed.","section":"References"},{"comment":"The caption contains the stray phrase 'a given image input to a Birds eye view', which appears to be a copy-editing artifact.","section":"Fig. 5 caption"},{"comment":"The vanishing-point training dataset is described only by source names and image counts; the paper does not explain how ground-truth vanishing points were obtained for the 172,576 images.","section":"§3.4"},{"comment":"The manual calibration tool is not fully specified: the paper does not state how the user chooses the z and x values, whether the tool provides visual feedback, or how sensitive the final BEV localisation is to those values.","section":"§4.2"}],"recommendation":"reject","confidential_remarks":"The manuscript's core claim appears to rest on a geometric construction that cannot yield metric measurements without additional calibration, and the evaluation is largely qualitative and circular. The nuScenes table raises reproducibility concerns because no protocol is given. I would advise the editor that, even if the ideas were developed further, the current manuscript would require a major reframing and substantial new experimentation before it could be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe headline: this paper puts together a genuinely new combination—learned vanishing point, YOLO/DeepSORT, and a homography built from four image points—to produce vector BEV maps from uncalibrated CCTV/webcam imagery. The qualitative overlays are visually convincing, and the author clearly put effort into a diverse VP training set (172k images) and an explainable pseudo-3D bounding box algorithm. If the paper were scoped as a qualitative urban-analytics tool (visualisation, anonymised token streams, occupancy patterns), it would be a reasonable contribution.\n\nThe problem is the central claim: 'high validation in geolocating road users' in metric terms. The stress-test lands. A single vertical vanishing point fixes the horizon but not focal length, camera height, or roll. The homography is constructed from two horizontal scanlines and two radial lines through the VP, mapped to an arbitrary rectangle; that gives a projective transformation, not a metric one. The subsequent manual calibration adjusts only a global scale (z) and an x-shift, which cannot compensate for anisotropic stretch or shear between the BEV axes. So the two-metre social-distancing analysis and the claimed GPS-level localisation are not entailed by the method. The Google Maps overlays are shown after manual alignment, so they cannot validate the metric claim independently.\n\nThe quantitative comparison in Table 4 makes things worse: there is no description of how TopView was applied to nuScenes, no code, and no evaluation protocol. Those numbers look fabricated. Table 2's VP MSE is fine, but it doesn't validate the downstream geometry.\n\nMy recommendation: reject the paper as it stands, but do not desk-reject without a careful technical review. The core idea is worth an expert referee's time to explain the geometric limitations and to require a proper metric evaluation (e.g., known camera parameters, or a direct comparison of BEV distances against ground truth). If the author reframes the claims to qualitative visualisation or adds a real calibration estimation stage, a revised version could be salvageable.\n\nI would not cite it in its current form, though I'd bring it to a reading group as a cautionary example of why vanishing points alone don't give metric homographies.\n\nBest,","headline":"A genuinely new pipeline for vector BEV maps from uncalibrated imagery, but the metric geolocation claim doesn't survive the geometry and the quantitative comparison is unexplained.","tokens_in":16351,"tokens_out":3962,"would_cite":false,"duration_ms":38434,"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":"This paper claims that a vector bird's-eye map of road users can be generated from a single uncalibrated street-level image or video by learning the scene's vanishing point, without knowing camera intrinsics or extrinsics.","keywords":["bird's eye view","homography","deep learning","urban scenes","vanishing point","object detection","uncalibrated camera","trajectory mapping"],"falsifier":"Take a calibrated camera on a flat, level ground plane with known intrinsics and extrinsics, photograph or render a scene, run TopView, and compare the BEV coordinates of objects against the ground-truth metric positions from the calibration. Systematically larger errors for objects farther from the bottom-center of the frame, or for scenes with a visible camera roll such as a tilted horizon, would falsify the claim that a vanishing point plus two scalars is sufficient for metric BEV reconstruction.","tokens_in":15420,"feed_emoji":"🗺️","tokens_out":5177,"duration_ms":42259,"temperature":0.7,"pith_summary":"This paper claims that a vector bird's-eye map of road users can be generated from a single uncalibrated street-level image or video by learning the scene's vanishing point, without knowing the camera's intrinsic or extrinsic parameters. The core idea is to regress the vanishing point with a deep network and then use that point to construct a homography that projects detected road users onto a top-down plane. The paper also derives pseudo-3D bounding boxes and orientation labels deterministically from 2D boxes, the vanishing point, and trajectory lines. If correct, any CCTV feed or internet image becomes convertible into a metric, map-aligned representation of traffic and pedestrians, enabling distance measurement, conflict detection, and city-scale occupancy analytics without per-camera calibration.","feed_headline":"One vanishing point maps any street camera to a top view","feed_subtitle":"A deep network learns the vanishing point, then homography projects road users onto a top-down map without camera calibration.","key_machinery":"The central object is the learned vanishing point of the scene, combined with a homography built from it. The vanishing point is regressed by a MobileNet-based network trained on 172,576 images from six datasets. The homography module draws a horizontal line near the bottom of the image, subdivides it uniformly, and connects those subdivisions to the vanishing point; the intersections of these radial lines with upper and lower horizontal bounds define a four-point quadrilateral that is mapped to a BEV rectangle. Algorithm 1 then converts a 2D box plus trajectory line into a 3D box by comparing the trajectory line's intersection with the box's top edge against the midpoint and the vanishing point's offset from image center, yielding orientation tags such as 'turning left', 'moving straight', or 'side view'.","core_discovery":"The central claim is that an orthogonal (bird's-eye) projection of road users can be estimated purely from the geometry of a single uncalibrated image by learning its vanishing point. TopView regresses the vanishing point and horizon from the image, detects road users with a YOLO detector, and then constructs a quadrilateral in the image plane—formed by radial lines from the vanishing point to a horizontal line at the bottom of the frame—that is mapped by homography to a rectangle in the BEV vector space. The same vanishing point, together with an object's trajectory line, constrains a 3D bounding box inside each 2D box via a deterministic geometric algorithm. The paper argues that this procedure geolocates road users accurately on map data (e.g., Google Maps) across diverse uncalibrated CCTV feeds, and reports competitive 3D detection metrics on the nuScenes validation set while requiring no calibration.","pith_inferences":["The vanishing-point-plus-two-scalars recipe is essentially a degenerate camera model with unknown focal length and height absorbed into two parameters; if the geometry holds, it should extend naturally to calibrated settings by replacing those scalars with true intrinsics and extrinsics, giving a fully parameter-free BEV for any camera.","A testable extension is to apply the same homography construction to scenes with visible vertical structures (buildings, poles) by using the learned vanishing point to also constrain scale along the vertical axis, which the current pipeline ignores.","The claimed city-scale social-distancing analysis implicitly assumes that the two manual scalars generalize across all cameras in a network; an explicit per-camera error analysis on a held-out set would be a natural next step.","The trajectory-based 3D box algorithm only works for moving objects; a direct extension for stationary objects would require inferring orientation from the object's pose relative to the vanishing point, which the paper leaves open."],"forward_implications":["Any uncalibrated CCTV feed or internet photo can be converted into a vector BEV map of road users, enabling distance and trajectory measurement without per-camera calibration.","Road-user positions can be localized on a geographic map using only the vanishing point and two manual scalars, supporting city-scale analyses of traffic and pedestrian exposure.","Videos can be reduced to anonymous spatiotemporal token streams (object id, class, 3D box, stationary status), which can be stored and retrieved compactly while preserving privacy.","The same framework enables analysis of social distancing violations across hundreds of cameras at city scale, as demonstrated on 857 London CCTV feeds.","The reported nuScenes metrics suggest that the method is competitive with camera-only 3D detection methods despite requiring no calibration, implying that the geometric shortcut may be a viable substitute for learned depth in some settings."],"supporting_citations":[{"why":"Supplies the street-view training images (London, Boston, Norway) and the vanishing-point ground-truth labels used to train the VP regression model.","marker":"[57]"},{"why":"Supplies additional AVA and Flickr images with vanishing-point annotations, expanding the diversity of the training set.","marker":"[58]"},{"why":"Provides the MobileNet backbone architecture for the vanishing-point regression network.","marker":"[55]"},{"why":"Provides the YOLOv5m object detector used to localize road users in each frame.","marker":"[23]"},{"why":"Provides the COCO dataset on which the detector is pre-trained, enabling detection of pedestrians, cars, buses, trucks, bicycles, and motorbikes.","marker":"[47]"},{"why":"Provides the DeepSORT tracker that maintains object identities across frames, crucial for trajectory extraction and temporal tokenization.","marker":"[48]"},{"why":"Provides a geometric approach for obtaining a bird's-eye view from a single image, which the homography module builds upon and extends to uncalibrated settings.","marker":"[32]"},{"why":"Supplies the TfL London CCTV feeds used to demonstrate and validate geolocalisation and city-scale applications.","marker":"[59]"},{"why":"Provides the nuScenes benchmark used for the quantitative 3D object detection comparison in Table 4.","marker":"[69]"}],"fun_headline_variants":["Deep learning maps street views without camera calibration","One vanishing point turns any camera into a top-down map","Bird's-eye view from uncalibrated street cams, no setup","Learn the vanishing point, skip the calibration step entirely"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole bird's-eye geometry rests on the assumption that a single learned vanishing point, plus two hand-set scalars for depth and sideways offset, fully determines a metric top-down view of an arbitrary street scene, which breaks if the ground is not flat or the camera is rolled relative to the scene.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning maps street views without camera calibration","One vanishing point turns any camera into a top-down map","Bird's-eye view from uncalibrated street cams, no setup","Learn the vanishing point, skip the calibration step entirely"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1198,"prompt_tokens":935,"completion_tokens":263,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":196}},"tokens_in":551,"tokens_out":263,"duration_ms":3203,"temperature":1.0,"reasoning_tokens":196,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T12:18:21.609174+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a calibrated camera on a flat, level ground plane with known intrinsics and extrinsics, photograph or render a scene, run TopView, and compare the BEV coordinates of objects against the ground-truth metric positions from the calibration. Systematically larger errors for objects farther from the bottom-center of the frame, or for scenes with a visible camera roll such as a tilted horizon, would falsify the claim that a vanishing point plus two scalars is sufficient for metric BEV reconstruction.","supporting_citations":[{"cited_title":"NeurVPS: Neural Vanishing Point Scanning via Conic Convolution","cited_arxiv_id":"1910.06316","evidence_quote":"Supplies additional AVA and Flickr images with vanishing-point annotations, expanding the diversity of the training set."},{"cited_title":"https://github.com/ultralytics/yolov5 (2021)","cited_arxiv_id":null,"evidence_quote":"Provides the YOLOv5m object detector used to localize road users in each frame."},{"cited_title":"https://github.com/nwojke/deepsort.git (2019)","cited_arxiv_id":null,"evidence_quote":"Provides the DeepSORT tracker that maintains object identities across frames, crucial for trajectory extraction and temporal tokenization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides a geometric approach for obtaining a bird's-eye view from a single image, which the homography module builds upon and extends to uncalibrated settings."},{"cited_title":"https://www.tfljamcams.net/ (2021)","cited_arxiv_id":null,"evidence_quote":"Supplies the TfL London CCTV feeds used to demonstrate and validate geolocalisation and city-scale applications."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020)","cited_arxiv_id":null,"evidence_quote":"Provides the nuScenes benchmark used for the quantitative 3D object detection comparison in Table 4."}],"review_version":1}