BIMCaP: BIM-based AI-supported LiDAR-Camera Pose Refinement

Andr\'e Borrmann; Anna Ribic; Borja Garc\'ia de Soto; Miguel Arturo Vega Torres

REVIEW 4 major objections 4 minor 3 references

BIMCaP: BIM-based AI-supported LiDAR-Camera Pose Refinement

T0 review · 4 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read BIMCaP refines drifted SLAM poses against a building information model and reports a translation-error reduction of over 4 cm compared with the state-of-the-art FACaP method.

desk verdict BIMCaP is a useful integration for BIM-based pose refinement, but its headline 4 cm translation claim rests on a single synthetic run with no error bars; the rotation improvements are more consistent and the paper deserves a revise-and-resubmit. read the letter →

arxiv 2412.03434 v1 pith:NOMLLKJX submitted 2024-12-04 cs.RO cs.AI

Miguel Arturo Vega Torres , Anna Ribic , Borja García de Soto , André Borrmann This is my paper · ORCID

classification cs.ROcs.AI

keywords buildinginformationmodelbundleadjustmentSLAMdriftcorrectionLiDAR-camerafusionsemanticsegmentationposerefinementindoormappingconstructionsite

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces BIMCaP, a pipeline that aligns mobile RGB and sparse-LiDAR data with a pre-existing building information model (BIM) to correct the drifted sensor poses typical of SLAM in indoor construction sites. Its central claim is that a bundle adjustment step, constrained by semantic alignment of walls, floors, columns, and the ceiling to the BIM, lowers the absolute trajectory error by more than 4 cm relative to the FACaP baseline on the ConSLAM dataset. This matters because it offers a route to accurate indoor 3D mapping with inexpensive sensors and without GPS, supporting construction monitoring and emergency response.

What carries the argument

The central mechanism is a semantic-aware bundle adjustment with a geometric term from multi-view feature correspondences between overlapping images and model-alignment terms: a floor term pulling segmented floor points to the BIM floor plane, a ceiling term, and wall and column terms that correct yaw and horizontal translation. These are supported by two earlier components: a hybrid depth-completion step combining linear interpolation with CompletionFormer to densify sparse LiDAR projections, and a modified Grounding DINO/RTMDet semantic segmentation pipeline that labels permanent elements. The BIM is preprocessed into a 2D vectorized semantic floor plan by projecting uniformly sampled synthetic point clouds and extracting line segments.

What would settle it

Run BIMCaP on sensor data from the ConSLAM dataset using a SLAM method that succeeds on that site (for example, a LiDAR-inertial odometry system) to generate real drifted poses, then compare the ATE after refinement. If the translation improvement over the unrefined trajectories is substantially less than 4 cm or the rotation error worsens, the synthetic-drift evaluation overstates the method's practical benefit.

Watch

Extended reading notes

Core claim

BIMCaP demonstrates that permanent building elements extracted from real-world images by semantic segmentation can serve as reliable anchors for pose correction: after projecting semantic labels into 3D with completed depth maps, a bundle adjustment module with cost terms for floor, ceiling, walls, columns, and multi-view geometric consistency aligns the measured map to a vectorized BIM floor plan. In the main experiment, the method reduces the synthetic-drift translation ATE from 1.391 m to 1.281 m, which is 4 cm lower than the FACaP result of 1.321 m, while also improving yaw and pitch errors in the smaller-drift experiments. The paper also shows that using only the LiDAR points inside the camera field of view keeps the method applicable to low-cost, reduced-FoV sensors.

Load-bearing premise

The paper assumes that its synthetic random-walk trajectories with increasing per-step error faithfully represent the drift that real SLAM systems would produce on the ConSLAM site; if real drift shows different patterns, the measured 4 cm gain and rotation improvements may not transfer to actual SLAM-based mapping.

Editorial extensions

If this is right

If the claimed accuracy holds, SLAM trajectories drifted by roughly 1.4 m can be pulled back to within about 1.28 m of ground truth using only a BIM and low-cost sensors.
The method does not require the full LiDAR field of view, so it transfers to solid-state LiDAR and RGB-D cameras that sense only part of the scene.
Construction site managers and emergency responders could obtain up-to-date, model-aligned digital maps without terrestrial laser scanning.
Since the floor and ceiling terms use the BIM planes instead of fitting planes to measurements, the optimized map aligns to the model's reference surfaces, avoiding bias from cluttered or uneven real surfaces.
The modular cost terms allow ablation; experiments show the combination of all five terms yields the best translation accuracy, though rotation refinement is not always best with all terms.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The reported 4 cm gain is only validated against synthetic random-walk drift, not against outputs of a working SLAM system, because DROID-SLAM and Go-SLAM failed on ConSLAM; a natural next test is to run BIMCaP on trajectories from a SLAM that does succeed on the same site and measure whether the gain persists under realistic drift statistics.
Experiment 3, where a pure rotation offset produced unintended translations, suggests the optimizer couples translation and rotation; a staged or regularized optimization may be needed to prevent cross-talk between the cost terms.
Because the BIM is reduced to a 2D floor plan for alignment, vertical (z-axis) drift may be only weakly observable; the method's strength is likely in horizontal translation and yaw correction, so users with strong elevation drift should expect less correction.
The depth completion quality is assessed only indirectly through pose metrics; a direct evaluation of the completed depth against ground-truth depth would isolate the contribution of the fusion step.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: BIMCaP does a sensible job of combining existing pieces—BIM reference surfaces, sparse LiDAR-camera depth fusion, and bundle adjustment—into a pipeline that works on a public construction-site dataset. The genuinely new bit is the specific set of semantic anchors (floor, ceiling, walls, columns) and the depth-completion step. The paper is honest about its failures, which I respect: Exp. 1 hits a local minimum, and Exp. 3 creates an unintended translation error from a zero-translation start. The rotation improvements are consistent across experiments, and the comparison against FACaP on ConSLAM is a fair baseline.

The soft spots are where the evidence gets shaky. The headline 'over 4 cm' translation gain comes from a single realization of a synthetic random-walk drift in Exp. 1. Exp. 2 shows BIMCaP making translation slightly worse (30.3 to 30.4 cm), and Exp. 3 introduces a 7.2 cm error from zero. No repeats, no error bars. Since the input drift is stochastic, the 4 cm gap in Exp. 1 could easily be within run-to-run variation. On top of that, the paper asserts the random-walk model matches 'usual trajectory patterns' of SLAM without validating it against any real SLAM output; the authors state that DROID-SLAM and Go-SLAM failed on ConSLAM, so the entire evaluation depends on synthetic poses. That is not fatal—the method is meant for small drift—but it does not support the abstract's global 'superior accuracy' claim. Finally, the semantic segmenter was trained on 250 labeled ConSLAM images and then used on the same dataset without a stated held-out split, which could inflate segmentation quality and, in turn, the pose-refinement results.

Who should read this: researchers in construction robotics, BIM-SLAM, and indoor mapping with low-cost sensors. The integration is a useful data point, and the public dataset helps reproducibility. The paper deserves a serious referee, but it needs substantive revisions: multiple synthetic trajectories with error bars, a held-out split for the segmenter, validation against real SLAM trajectories where possible, and a more measured abstract. I would not desk-reject it, but I would not accept it as is either.

My recommendation: send it to peer review with the expectation of a major revision.

Referee Report

4 major / 4 minor

Summary. The paper proposes BIMCaP, a method for refining camera and sparse-LiDAR sensor poses against a pre-existing building information model (BIM). The pipeline consists of (1) depth completion from sparse LiDAR and RGB images, (2) semantic segmentation of permanent elements (walls, floor, columns, ceiling) and construction of a vectorized semantic floor plan from the BIM, and (3) a bundle-adjustment optimization with geometric, floor, ceiling, wall, and column cost terms. Experiments on the ConSLAM dataset compare BIMCaP with the FACaP baseline under three synthetic-drift scenarios, reporting a 4 cm reduction in translational ATE in the first experiment. The paper claims that BIMCaP achieves superior accuracy over current state-of-the-art methods and is applicable to indoor mapping for construction and emergency response.

Significance. If the reported accuracy improvement is robust, BIMCaP would be a useful contribution to aligning low-cost SLAM outputs with BIMs in cluttered indoor environments. The paper uses an open-access dataset (ConSLAM), provides a public repository link, and performs an ablation of the different cost terms, which are strengths for reproducibility and for understanding the method's components. However, the central quantitative claim of 'over 4 cm' translation improvement rests on a single synthetic trajectory without error bars, and the evaluation is weakened by the lack of validation of the synthetic drift model and by the apparent overlap between the training data for the semantic segmenter and the evaluation data. These issues must be addressed before the claimed superiority is established.

major comments (4)

[§4.2, Table 1] The headline claim of 'reducing translational error by over 4 cm' is supported by only one experimental condition (Exp. 1) and one stochastic realization: ATEpos drops from 1.321 m (FACaP) to 1.281 m (BIMCaP), a 0.040 m gap. No repeated runs, seeds, or error bars are reported, despite the input poses being sampled from a random-walk process (Section 3.3.1). The paper should report the variance of the ATE metrics over multiple independent synthetic-trajectory draws, and the abstract/conclusions should be tempered unless the 4 cm advantage is shown to be statistically distinguishable from run-to-run variation.
[§3.3.1, §4.2] The synthetic trajectories used in all three experiments are generated by a normally distributed random-walk drift model, which is asserted to match 'the usual trajectory patterns seen in existing SLAM systems' but is not validated against any real SLAM output. The authors state that DROID-SLAM and Go-SLAM failed on ConSLAM, so the synthetic model is not calibrated against any real drift statistics. Without such validation, it is unclear whether the reported improvements transfer to actual SLAM-based mapping. The paper should either validate the drift model on a dataset where a SLAM algorithm succeeds, compare the synthetic drift statistics with real SLAM drift, or clearly limit the claims to the synthetic drift regime.
[§3.2.2, §4.1] The RTMDet object detector used for semantic segmentation was trained on 250 labeled images from the ConSLAM dataset, and the same dataset is used for evaluating pose refinement. The paper does not describe any held-out split or cross-validation between the images used for training the segmenter and those used to generate the semantic point clouds for pose optimization. If the same images appear in both stages, the measured segmentation accuracy and the resulting pose-refinement results are likely optimistically biased. The authors should clarify the exact split (e.g., by sequence, by timestamp) or retrain/evaluate in a way that prevents leakage.
[§4.2, Table 2, Exp. 3] In Exp. 3, starting from zero translation error, BIMCaP introduces a 7.2 cm translation error, and the paper acknowledges 'unintended translations.' This result directly undermines the general claim that BIMCaP reduces translational error relative to the initial poses, and it also indicates that the optimization can create errors when the true drift is rotation-only. The paper should either restrict the translation-accuracy claim to the tested drift regimes, add a constraint or a separate handling for rotation-only cases, or explain why the translation error in Exp. 3 does not affect the headline claim.

minor comments (4)

[§3.3.2] The weights of the cost-function terms (geometric, floor, ceiling, wall, column) are not specified in the text, which makes the experiments hard to reproduce from the paper alone; please provide the weight values or reference where they are defined in the repository.
[§4.1] The metrics MME, MPV, and NND are described briefly, but the description of MME as 'higher is better' with the statement that the ground-truth value 0.761 'represents the optimal alignment' is confusing because MME is not bounded above; a short definition of each metric would improve clarity.
[Throughout] There are minor inconsistencies in terminology, e.g., 'FACaP' vs. 'FaCAP' is used interchangeably, and the paper refers to 'Fig. 5b and 5c' while the figure itself has multiple subfigures; please standardize the naming and references.
[§1] The phrase 'We demonstrated the improvement in performance' appears in the introduction before any experimental evidence is presented; consider using a more neutral phrasing such as 'We evaluate the improvement...'.

Circularity Check

2 steps flagged · score 4.0 of 10

The pose-optimization core is not definitionally circular, but the headline 4 cm claim rests on a self-generated benchmark and a semantic segmenter trained on the same evaluation dataset.

fitted input called prediction [Section 3.2.2 (Semantic segmentation of real-world data) and Section 4.1 (Dataset and evaluation details)]
"we replace the DINO algorithm with a tiny version of the RTMDet algorithm (Lyu et al. 2022) pre-trained with the COCO dataset and 250 labeled images of the ConSLAM dataset ... To ensure reproducibility and enable benchmarking, we tested the developed method on the ConSLAM dataset (Trzeciak et al. 2023)."

The RTMDet-based segmenter is fitted to 250 labeled ConSLAM images and then used to produce the semantic labels that feed the bundle-adjustment cost functions. The paper reports no held-out split, so the same ConSLAM dataset is used both to train this component and to evaluate the full pipeline. The pose-refinement results, including the 4 cm ATE improvement over FACaP, are therefore measured on images the segmenter has already seen; the evaluation does not establish generalization, and part of the measured advantage can reflect the segmenter's fit to the evaluation data rather than an independent property of the refinement method.
self citation load bearing [Section 4.1 (Dataset and evaluation details)]
"The GT poses of ConSLAM were calculated using SLAM2REF (Vega-Torres et al. 2024), an enhanced version of BIM-SLAM (Vega Torres et al. 2023) and OGM2PGBM (Vega Torres et al. 2022) for large-scale maps, which is robust to LiDAR motion distortion and Scan-Map deviations."

All three cited methods share authors with the present paper, and SLAM2REF is itself a reference-map-based pose-refinement method. The benchmark ground truth against which BIMCaP's ATE improvement is measured is therefore produced by the same research line that is being evaluated. The paper provides no independent validation of these GT poses in this work, so the central quantitative claim (4 cm lower translational error than FACaP) rests on a self-citation chain: the yardstick is the authors' own prior method rather than an externally verified benchmark.

full rationale

The pose-refinement derivation itself is not circular: the bundle-adjustment cost functions (geometric, floor, ceiling, wall, column) are optimized against a BIM reference, and the reported ATE values are external measurements relative to ground-truth poses rather than algebraic consequences of the cost terms. However, two evaluation assumptions are self-referential enough to raise the score. First, the RTMDet-based semantic segmenter is trained on 250 ConSLAM images and then applied within the same ConSLAM evaluation with no stated held-out split, so the semantic inputs to the optimization are partly fitted to the evaluation data. Second, the ground-truth poses used for ATE were calculated by SLAM2REF, a prior method by the same authors that also aligns LiDAR to a reference map, and no independent validation of those poses is presented in the paper. These issues do not make the optimization equivalent to its inputs by construction, but they mean the headline 'reducing translational error by over 4 cm' is not fully independent of the authors' own fitted pipeline and prior work. The synthetic random-walk drift is a robustness concern rather than a circularity concern, and the comparison to FACaP provides some independent content, so a moderate score of 4 is appropriate.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method leans on several stated and unstated assumptions. The most consequential are the accuracy of the BIM, the representativeness of synthetic drift, and the quality of semantic segmentation trained on the evaluation dataset. No new physical entities are introduced.

free parameters (3)

Cost function term weights
The bundle adjustment cost combines geometric, floor, ceiling, wall, and column terms. The weights are not reported, and Table 1 shows results vary strongly with term selection, so these weights are tuned degrees of freedom in the central claim.
Synthetic trajectory hyperparameters (sigma_t, sigma_p, sigma_th) = not reported
Section 3.3.1 defines the random-walk drift variances as adjustable. The experiments' initial offsets (1.4 m/10 deg, 30.3 cm/8.82 deg, 0/9.6 deg) depend on these values, but the values themselves are not listed, so the evaluation conditions are not fully specified.
Reference-map creation thresholds
The 2D vectorized floor plan is generated by sampling the BIM, projecting within +/-20 cm of floor level, and applying line detection. Sampling density and detection thresholds are chosen by hand and not reported, yet they determine the reference map used for alignment.

assumptions (5)

domain assumption The BIM provides a geometrically accurate representation of the permanent environment, with centimeter-level fidelity.
ConSLAM's BIM is generated from TLS scans, and all alignment terms exploit it as a reference. Errors in the BIM become errors in the refined poses, as noted in Section 4.1.
domain assumption Walls and columns are perpendicular to the XY plane, so projecting the BIM to a 2D floor plan preserves the information needed for pose correction.
Section 3.2.1 relies on this geometric simplification for the vectorized semantic reference map.
domain assumption An initial rough alignment between the drifted trajectory and the BIM exists.
Stated at the end of Section 1. The method refines small drift and cannot recover from a large initial misalignment.
ad hoc to paper A random-walk synthetic pose sequence is representative of real SLAM drift.
Section 3.3.1 creates synthetic trajectories because DROID-SLAM and Go-SLAM fail on ConSLAM. The assertion that this matches 'usual trajectory patterns' is made without comparison to real SLAM outputs.
domain assumption Semantic segmentation of real images into walls, floor, columns, and ceiling is sufficiently accurate.
Step 2 (Section 3.2.2) depends on these labels to select reliable landmarks. Segmentation errors propagate into the optimized poses.

how reviews work

0 comments

Cite this review

Pith. "Pith review of BIMCaP: BIM-based AI-supported LiDAR-Camera Pose Refinement." pith.science (2026). https://pith.science/paper/NOMLLKJX

@misc{pith2026241203434,
  author       = {Pith},
  title        = {Pith review of: BIMCaP: BIM-based AI-supported LiDAR-Camera Pose Refinement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NOMLLKJX}},
  note         = {Machine review of arXiv:2412.03434}
}

read the original abstract

This paper introduces BIMCaP, a novel method to integrate mobile 3D sparse LiDAR data and camera measurements with pre-existing building information models (BIMs), enhancing fast and accurate indoor mapping with affordable sensors. BIMCaP refines sensor poses by leveraging a 3D BIM and employing a bundle adjustment technique to align real-world measurements with the model. Experiments using real-world open-access data show that BIMCaP achieves superior accuracy, reducing translational error by over 4 cm compared to current state-of-the-art methods. This advancement enhances the accuracy and cost-effectiveness of 3D mapping methodologies like SLAM. BIMCaP's improvements benefit various fields, including construction site management and emergency response, by providing up-to-date, aligned digital maps for better decision-making and productivity. Link to the repository: https://github.com/MigVega/BIMCaP

Figures

Figures reproduced from arXiv: 2412.03434 by the authors.

**Figure 2.** Depth completion with sparse LiDAR point cloud: (a) original image from the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Reference map preparation: (a) original 3D BIM (without ceiling); (b) uniformly [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Semantic segmentation over 2D images of the ConSLAM dataset: (a) inference [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Features used for optimization (a) Top view semantic segmented map generated [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Development of the translational (a) and rotational error (b) given only a rotational [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: Side views of the different maps. (a) ground truth map; (b) map created with synthetic [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages

[1]

& Bab-Hadiashar, A

Acharya, D., Tennakoon, R., Muthu, S., Khoshelham, K., Hoseinnezhad, R. & Bab-Hadiashar, A. (2022), ‘Single-image localisation using 3D models: Combining hierarchical edge maps and semantic segmentation for domain adaptation’,Automation in Construction 136, 104152. Boniardi, F., Valada, A., Mohan, R., Caselitz, T. & Burgard, W. (2019), ‘Robot localization...

work page 2022
[13]

& Borrmann, A

Vega Torres, M., Braun, A. & Borrmann, A. (2022), Occupancy grid map to pose graph-based map: Robust BIM-based 2D-lidar localization for lifelong indoor navigation in changing and dynamic environments, in S. F. S. Eilif Hjelseth & R. Scherer, eds, ‘eWork and eBusiness in Architecture, Engineering and Construction: ECPPM 2022’, CRC Press, Trondheim, Norway...

work page 2022
[33]

Kropp, C., Koch, C

Forum Bauinformatik’. Kropp, C., Koch, C. & K¨onig, M. (2018), ‘Interior construction state recognition with 4d BIM registered image sequences’, Automation in Construction 86, 11–32. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J. et al. (2023), ‘Grounding dino: Marrying dino with grounded pre-training for open-s...

arXiv 2018

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

& Bab-Hadiashar, A

Acharya, D., Tennakoon, R., Muthu, S., Khoshelham, K., Hoseinnezhad, R. & Bab-Hadiashar, A. (2022), ‘Single-image localisation using 3D models: Combining hierarchical edge maps and semantic segmentation for domain adaptation’,Automation in Construction 136, 104152. Boniardi, F., Valada, A., Mohan, R., Caselitz, T. & Burgard, W. (2019), ‘Robot localization...

work page 2022

[2] [13]

& Borrmann, A

Vega Torres, M., Braun, A. & Borrmann, A. (2022), Occupancy grid map to pose graph-based map: Robust BIM-based 2D-lidar localization for lifelong indoor navigation in changing and dynamic environments, in S. F. S. Eilif Hjelseth & R. Scherer, eds, ‘eWork and eBusiness in Architecture, Engineering and Construction: ECPPM 2022’, CRC Press, Trondheim, Norway...

work page 2022

[3] [33]

Kropp, C., Koch, C

Forum Bauinformatik’. Kropp, C., Koch, C. & K¨onig, M. (2018), ‘Interior construction state recognition with 4d BIM registered image sequences’, Automation in Construction 86, 11–32. Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Li, C., Yang, J., Su, H., Zhu, J. et al. (2023), ‘Grounding dino: Marrying dino with grounded pre-training for open-s...

arXiv 2018