{"id":"c2a3ac21-67ab-4e0c-8e13-c50ff4846ad6","arxiv_id":"2507.17445","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"IndoorBEV uses a query-based transformer decoder on a bird's-eye view lidar grid to jointly detect objects and predict footprint masks in indoor scenes.","lead":"IndoorBEV converts raw lidar point clouds from a mobile robot into a top-down bird's-eye view and predicts each object's class, position, and a pixel mask of its footprint. The authors claim this captures arbitrary object shapes, but their ground-truth masks are rasterized from 3D bounding boxes, so that claim is not actually tested.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'arbitrary-shape footprint' claim is untestable as written: every ground-truth mask in §4.1 is a rasterized oriented rectangle from a 3D box, so the model is neither trained nor evaluated on non-rectangular footprints.","rationale":"The reader's weakest assumption exactly identifies the load-bearing issue. The whole point of IndoorBEV is to go beyond bounding-box regression, but the annotation pipeline described in §4.1 reduces every object to an oriented rectangle. That means the training loss and the evaluation metrics can never distinguish a mask-prediction model from a bounding-box model; any apparent 'shape faithfulness' in Figs. 1 and 5 is almost certainly faithfulness to a rectangle. I looked for other soft spots: the lack of baselines and error bars is real, and the unreleased dataset and code weaken reproducibility, but those are secondary. The mask-label issue is more fundamental because it attacks the claimed contribution rather than only the strength of the evaluation. The appropriate disposition is unchanged from the reader's CONDITIONAL: the architecture is plausible and internally consistent, but before acceptance the authors should either add free-form mask annotations and a rectangle baseline or explicitly narrow the claim to rectangular-footprint completion. No ad hominem intended; this is a data/claim mismatch.","tokens_in":10228,"tokens_out":5789,"duration_ms":64233,"concrete_test":"Retrain IndoorBEV with the same architecture but supervise the mask head with free-form ground-truth masks (hand-drawn polygons, or masks directly rasterized from the MuJoCo primitive geometries for spheres, cylinders, capsules, and ellipsoids) instead of the §4.1 oriented-rectangle rasterization. Evaluate on held-out scenes containing non-rectangular objects: if AP@0.50 and mIoU with free-form masks do not exceed the scores obtained when the same predictions are scored against oriented-rectangle rasterizations, then the claimed advantage over bounding-box regression is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 states that labels are rasterized into 2D instance masks by a MaskRasterizer using '3D dimensions, location, and yaw' and mapping the object footprint to the BEV grid as a filled polygon. For a KITTI-style 3D box annotation, this polygon is an oriented rectangle. The paper never describes any free-form or geometry-derived mask annotation, so both simulated and real training and test masks are rectangles. The Abstract and §3 claim the mask-centric formulation captures footprints 'regardless of their shape' and that the shape prior comes from the ground-truth masks, but the actual shape prior contains only boxes. Consequently, the reported metrics in §4.6 (78.4% AP@0.25, 63.7% AP@0.50, 67.2% mIoU, 64.5% PQ) measure rectangle-mask accuracy, not arbitrary-shape completion. A bounding-box or oriented-rectangle detector could plausibly match these numbers because the supervision signal is equivalent. This is a genuine gap between the central novelty claim and the evidence, and it is not disclosed in §6 Limitations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IndoorBEV, a lidar-based bird's-eye-view (BEV) perception framework for indoor mobile robots. The method projects raw 3D point clouds into a BEV grid via an axis-fusion encoder, processes the BEV feature map with a window-based Swin-style backbone, and uses a query-based transformer decoder to jointly predict object class, 3D attributes (position, dimensions, yaw), and an instance-level footprint mask. The authors claim that this mask-centric formulation captures object footprints 'regardless of their shape' and provides a robust alternative to bounding-box regression. They evaluate on a custom hybrid dataset of simulated and real-world indoor lidar scans, reporting AP@0.25 of 78.4%, AP@0.50 of 63.7%, mIoU of 67.2%, and PQ of 64.5% on the real-world test set.","tokens_in":10416,"tokens_out":4424,"duration_ms":52543,"significance":"If the central claim were fully supported, IndoorBEV would be a useful contribution to indoor robotic perception, since a unified detection-and-footprint module operating directly on lidar in BEV space could benefit navigation, collision checking, and planning. The manuscript also provides a concrete algorithm description, a pseudocode specification, and an analytic ray-tracing simulator in the appendix, which are helpful for reproducibility. However, the main novelty as stated—arbitrary-shape footprint completion—is not established by the experiments, because all ground-truth masks are rasterized from 3D bounding boxes. The lack of baselines and ablations further prevents the reader from assessing whether the mask-based head is actually superior to bounding-box regression. The idea is promising and the core architecture is sound, but the evidence currently does not support the paper's strongest claims.","major_comments":[{"comment":"The ground-truth masks used for both training and evaluation are generated by the MaskRasterizer from KITTI-style labels containing 3D dimensions, location, and yaw. Such labels produce filled oriented rectangles on the BEV grid. Since both the simulated and real-world annotations follow this format, every mask in the dataset is rectangular. Consequently, the claims in the Abstract, Section 3, and Section 5 that the method captures footprints 'regardless of their shape' or handles 'arbitrary shapes' are not supported by the evidence. The metrics in Section 4.6 measure rectangle-mask accuracy, not arbitrary-shape footprint completion. To substantiate the central novelty, the authors must either introduce non-rectangular ground-truth masks (e.g., from mesh geometry or manual polygon annotation) or substantially rephrase the claims to describe rectangle-derived mask prediction and completion.","section":"Section 4.1, \"Data Format and Preprocessing\""},{"comment":"The experimental section reports only absolute numbers for IndoorBEV and contains no comparisons with any baseline, nor any ablation of the main components. Without an oriented-bounding-box BEV detector or a mask-based BEV method adapted to indoor scenes, the statements 'offering a robust alternative to bounding box regression' (Abstract) and 'capture object shapes more faithfully than traditional bounding boxes' (Section 4.6, Qualitative) are assertions rather than demonstrated conclusions. Please add at least one bounding-box regression baseline and one ablation that replaces the mask head with a box head, so that the benefit of mask-based prediction can be directly measured.","section":"Section 4.6, \"Results\""},{"comment":"The term 'footprint completion' is used in the title, the Introduction, and Figure 2, but the paper does not define what 'completion' means beyond mask prediction. The mask head in Section 3.3 computes mask logits as the inner product of mask embeddings with projected BEV features; there is no explicit mechanism for inferring occluded or truncated object regions. The evaluation metrics in Section 4.5 (mask IoU, mIoU, PQ) only measure overlap with the full ground-truth mask and cannot distinguish completion of partially observed objects from simple mask fitting. Please provide a precise definition of footprint completion and add an experiment that evaluates it, for example by artificially truncating lidar points and measuring how much of the full footprint is recovered, or remove the completion terminology.","section":"Section 3.3 and Figure 2"},{"comment":"The paper repeatedly claims that IndoorBEV 'operates in real-time' and is suitable for mobile robot deployment, but no runtime or latency measurements are reported anywhere in Section 4. The only timing-related statement concerns the simulator in Section 3.4, not the inference of the perception model. Since real-time performance is presented as a key advantage for robotics, please report inference time (e.g., milliseconds per frame, or frames per second) on the target hardware, including the computational environment used for the experiments. Without these numbers, the real-time claim is unsupported.","section":"Introduction and Section 5"}],"minor_comments":[{"comment":"The text states that the mask-based decoder uses '256 attention heads' while the hidden dimension is Chid = 64. In a standard Transformer, the model dimension should equal the number of heads times the head dimension, so 256 heads with a 64-dimensional model is dimensionally inconsistent. Please clarify whether this is a typo (e.g., 4 or 8 heads) or whether a different attention construction is used.","section":"Section 4.4, \"Implementation Details\""},{"comment":"The backbone is described as modified from 'SwinTransformer [15]', but reference [15] is the Tutel paper on mixture-of-experts, not the Swin Transformer paper. Please cite the correct Swin Transformer source, and ensure all references are checked for consistency.","section":"Section 3.2 and Reference [15]"},{"comment":"The training setup is described as 'five nodes each equips with two Intel Xeon processors ... and contains four Nvidia A100 GPUs'. It is unclear whether this means five nodes with four GPUs per node (20 GPUs total) or five nodes sharing four GPUs. Please clarify the total number of GPUs used for training.","section":"Section 4.4, \"Implementation Details\""},{"comment":"The matching cost in Equation (2) uses coefficients αcls, αdice, αfoc, and αreg, while the final loss in Equation (3) uses λcls, λdice, λmask, λdim, λpos, and λyaw. The relationship between these two sets of weights is not stated. Please specify how the α values are set and how they relate to the λ weights listed in Section 4.3.","section":"Equations (2) and (3)"},{"comment":"There are numerous typos and grammatical issues, including 'handles naturally occlusions' in the Abstract, 'each secnes' in Figure 1 and Figure 5 captions, and 'the algorithm were trained' in Section 4.4. A careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a robotics conference and the core architecture is reasonable, but the experimental evaluation is currently too weak to support the central arbitrary-shape and footprint-completion claims. The revision should prioritize (1) non-rectangular ground-truth masks or a clear downgrading of the shape-generality claims, (2) at least one bounding-box baseline and an ablation of the mask head, and (3) actual runtime numbers for the real-time claim. I would not require the authors to collect an entirely new dataset if they rephrase the claims accurately, but without those changes the contribution as stated is not convincing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the quick read on IndoorBEV. The paper does something new in a narrow sense: it applies mask-based BEV detection to indoor lidar for mobile robots, using a cluster encoder, a Swin backbone, and a Mask2Former-style decoder. That combination hasn't appeared indoors before, and the MuJoCo-Taichi simulator they built is a genuinely useful piece of engineering. The writing is clear and the architecture is described well enough to reimplement.\n\nThe problem is the headline claim. The abstract and Section 3 say the mask-centric formulation captures footprints 'regardless of their shape.' But Section 4.1 says ground-truth masks are rasterized from 3D dimensions, location, and yaw via a MaskRasterizer that maps the footprint to a filled polygon. For KITTI-style box annotations, that polygon is an oriented rectangle. Every mask in the train and test sets is box-shaped. So the model never sees arbitrary or non-rectangular footprints, and the metrics in Section 4.6 measure rectangle-mask accuracy, not shape completion. A good bounding-box detector could plausibly match those numbers because the supervision is equivalent. The paper doesn't disclose this in the Limitations section.\n\nThere are other soft spots, in proportion: no baseline comparisons (MaskBEV is the obvious one), no ablations of the mask head, no error bars, and the real dataset is five indoor environments. Code and data are not released. These are all addressable, but together they mean the paper currently stands as a demo rather than a supported advance.\n\nI want to be fair: the architecture is sensible, the simulator is a real contribution, and the numbers, while not benchmarked, are plausible for a supervised detector. The paper is not a hoax or a toy; it just doesn't support its central novelty claim as written.\n\nWho gets value from this? People working on indoor robot perception, especially those wanting a concrete mask-based BEV baseline. They should read it for the system design and the simulator, and ignore the 'arbitrary shape' rhetoric.\n\nFor peer review: I'd send it to reviewers, but with a strong expectation of major revision. The gap between claim and evidence is fixable—swap or supplement the box-derived masks with free-form annotations, add at least MaskBEV and a box-regression baseline, release the simulator—and the core idea is worth testing properly.\n\nRegards.","headline":"A useful indoor BEV detection demo undermined by an unsupported arbitrary-shape claim—all ground-truth masks are box-derived rectangles.","tokens_in":10990,"tokens_out":2304,"would_cite":false,"duration_ms":24156,"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 claims that indoor lidar perception can be reframed as mask prediction in bird's-eye view, producing exact object footprints instead of bounding boxes.","keywords":["bird's-eye view","lidar point cloud","indoor robot perception","instance mask prediction","footprint completion","query-based decoder","object detection","mobile robots"],"falsifier":"Collect an evaluation set of indoor objects with genuinely irregular footprints (curved benches, L-shaped shelves, sprawled cables) whose BEV masks are hand-annotated rather than rasterized from 3D boxes, then measure IndoorBEV's mask IoU against the same model retrained with box-regression; if the mask model's advantage over boxes disappears on this set, the central claim of shape-agnostic footprint completion is not supported.","tokens_in":9987,"feed_emoji":"🤖","tokens_out":7823,"duration_ms":79544,"temperature":0.7,"pith_summary":"IndoorBEV aims to give an indoor mobile robot a single bird's-eye-view (BEV) perception module: given raw lidar point clouds, it outputs for every detected object its class, 3D pose attributes, and a dense footprint mask instead of a bounding box. The paper's rationale is that bounding boxes, which suit outdoor road vehicles, distort irregular indoor objects, whereas masks preserve the shape needed for collision avoidance and planning. It claims to be the first BEV perception framework designed specifically for indoor mobile robots and reports 78.4% AP@0.25, 63.7% AP@0.50, 67.2% mIoU, and 64.5% PQ on a real-world test set from its custom hybrid dataset. If the claim holds, robots could build top-down scene understanding in a form directly usable by navigation and planning stacks.","feed_headline":"Mask-based BEV detection maps indoor objects by exact footprint","feed_subtitle":"A lidar-only model returns each object's full top-down silhouette, giving indoor robots shape-aware input for navigation.","key_machinery":"The load-bearing component is the query-based mask decoder (SimpleMaskHead), which converts each refined object query into a mask embedding and computes mask logits as the dot product of that embedding with the projected BEV feature map, so every query specializes to one instance's footprint. This replaces bounding-box regression and is trained with set-based bipartite matching between queries and ground-truth instances, so the network learns masks directly instead of fitting boxes. A second mechanism is the axis-compact cluster encoder (AF-PCCN), which fuses X-Z and Y-Z axis features of lidar clusters and scatters them to the BEV grid, marking empty clusters as noise instead of padding them with zeros. The shape prior that makes footprint completion possible comes from ground-truth masks, rasterized as filled polygons by the MaskRasterizer utility.","core_discovery":"The central claim is that detection and footprint completion can be performed jointly as mask prediction in BEV space, and that this mask-centric formulation captures the footprint of static and dynamic objects regardless of shape, offering a robust alternative to bounding-box regression. The architecture projects clustered lidar points onto a BEV grid, enriches the grid with a windowed self-attention backbone, and lets learned object queries interact with the BEV features; each query then predicts a class, dimensions, position, yaw, and a per-pixel mask computed by multiplying the query's mask embedding with the projected feature map. The predicted masks are said to show full object edges even when the raw BEV projection is sparse, because the ground-truth mask annotations provide a shape prior. The paper presents this as the first BEV pipeline aimed at indoor mobile-robot perception, evaluated on a dataset that mixes simulated and real-world lidar frames.","pith_inferences":["In the paper's setup every training and evaluation mask is rasterized from a 3D bounding box, so the arbitrary-shape claim is verified only against rectangular footprints; the next test is to train on hand-annotated masks for curved or L-shaped objects to see whether the mask head truly generalizes beyond box priors.","The same encoder-decoder design could be reused for BEV panoptic mapping, where a temporal memory merges per-frame masks into a persistent top-down map without a separate clustering stage.","The axis-compact encoding that labels empty clusters as noise is a transferable idea for any sparse lidar processing, indoor or outdoor, and could be measured in isolation by ablating it against standard pillar padding.","Performance against a bounding-box baseline on the same architecture is not reported, so a controlled comparison would reveal how much of the stated accuracy comes from the mask head and how much from the encoder and backbone."],"forward_implications":["Each detected object arrives with a pixel-level footprint, so downstream costmaps and planners can reason about exact occupied extents rather than inflated boxes.","Static furniture and dynamic agents are handled by the same mask head, so a single model can feed both collision checking and motion prediction.","Because masks, classes, positions, and yaw are produced in one forward pass, the BEV output can be connected directly to navigation without a separate detection-to-map stage.","The claimed real-time operation implies the framework can support closed-loop onboard perception on a mobile robot.","The mask output supports future extensions to temporal tracking and instance-level reasoning, since each query consistently targets an object."],"supporting_citations":[{"why":"Establishes the precedent of reframing BEV detection as mask prediction, but only for outdoor urban scenes, which this paper extends to indoor robots.","marker":"[10]"},{"why":"Used to support the claim that ground-truth mask annotations supply the shape prior enabling completion of arbitrary object shapes.","marker":"[12]"},{"why":"Supplies the set-based, query-driven detection and bipartite matching training scheme that the mask decoder adapts.","marker":"[16]"},{"why":"Provides the mask-transformer design pattern the decoder head follows for per-query mask generation.","marker":"[17]"},{"why":"Provides the physics simulator used to generate procedurally varied synthetic indoor scenes for training.","marker":"[18]"},{"why":"Provides the parallel ray-casting approach used to render realistic lidar scans in the simulation pipeline.","marker":"[19]"},{"why":"Used in the semi-automatic annotation pipeline for generating and verifying real-world footprint masks.","marker":"[20]"},{"why":"Defines the data and label format (object type, dimensions, location, yaw) that the custom dataset stores frames in.","marker":"[21]"}],"fun_headline_variants":["Mask-based BEV gives indoor robots full object footprints","IndoorBEV: joint detection and footprint completion via masks","Lidar BEV masks replace boxes for indoor object detection","IndoorBEV predicts object shapes in bird's-eye view"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The system is trained and evaluated on masks that are generated by rasterizing 3D bounding boxes, so arbitrary-shape capture is only demonstrated for box-shaped footprints; if a robot meets genuinely non-rectangular annotated objects, the claimed advantage over bounding boxes is unproven.","fun_headline_variants_meta":{"raw":{"variants":["Mask-based BEV gives indoor robots full object footprints","IndoorBEV: joint detection and footprint completion via masks","Lidar BEV masks replace boxes for indoor object detection","IndoorBEV predicts object shapes in bird's-eye view"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1357,"prompt_tokens":967,"completion_tokens":390,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":321}},"tokens_in":583,"tokens_out":390,"duration_ms":4613,"temperature":1.0,"reasoning_tokens":321,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:47:15.380661+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect an evaluation set of indoor objects with genuinely irregular footprints (curved benches, L-shaped shelves, sprawled cables) whose BEV masks are hand-annotated rather than rasterized from 3D boxes, then measure IndoorBEV's mask IoU against the same model retrained with box-regression; if the mask model's advantage over boxes disappears on this set, the central claim of shape-agnostic footprint completion is not supported.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the precedent of reframing BEV detection as mask prediction, but only for outdoor urban scenes, which this paper extends to indoor robots."},{"cited_title":"PiMAE: Point Cloud and Image Interactive Masked Autoencoders for 3D Object Detection","cited_arxiv_id":"2303.08129","evidence_quote":"Used to support the claim that ground-truth mask annotations supply the shape prior enabling completion of arbitrary object shapes."},{"cited_title":"Carion, F","cited_arxiv_id":null,"evidence_quote":"Supplies the set-based, query-driven detection and bipartite matching training scheme that the mask decoder adapts."},{"cited_title":"Cheng, I","cited_arxiv_id":null,"evidence_quote":"Provides the mask-transformer design pattern the decoder head follows for per-query mask generation."},{"cited_title":"Hu, T.-M","cited_arxiv_id":null,"evidence_quote":"Provides the parallel ray-casting approach used to render realistic lidar scans in the simulation pipeline."}],"review_version":1}