CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction

Thank you for your constructive feedback. We will carefully revise the paper to address your concerns and sincerely appreciate your thoughtful review and critical insights.

Question 1: Contribution with respect to RoomFormer

Answer:

• CAGE introduces two key contributions: (1) an edge-based polygon representation that explicitly encodes geometry and direction, enhancing robustness to occlusion and missing corners (e.g., +7.7% Angle F1); and (2) a dual-query denoising transformer decoder, which extends DN-DETR to sequential polygon prediction via perturbed edge queries. Ablation results (Appendix B.5) show that removing either component leads to significant drops (–5.8% Corner F1, –1.9% Angle F1).
• This is far from a trivial extension. Our edge-centric formulation directly addresses core limitations of corner-based approaches and ensures topological consistency. As recognized by R#fJqK, "the edge representation encodes directionality explicitly and promotes angular regularity… beneficial in floorplan reconstruction." R#Sxuu called it a "simple but efficient change… maintaining directional and structural consistency." Besides, R#EK67 affirmed that "The dual-query decoder with denoising supervision is both novel and effective… robust under occlusion." Notably, while DN-DETR is an existing method, we are the first to adapt it for floorplan reconstruction task. By leveraging denoising supervision, we obviously enhance both accuracy and stability, making this a highly valuable contribution.
• Together, these innovations yield significant accuracy and robustness gains. We establish a foundation for broader raster-to-vector tasks, such as façade modeling and cadastral map reconstruction, where we are actively extending our framework.

Question 2: Hardcoded maximum number of rooms and edges

Answer:

• Our method is inherently scalable and built on a general framework, with the primary constraint being token length, which is limited by available GPU memory. We use a configuration of 800 tokens (40 polygons × 20 edges), which is sufficient for most residential layouts. For example, Structured3D typically contains fewer than 10 rooms per suit, and SceneCAD includes only one room per case. As demonstrated in case 3410 (Fig. 10, supplementary), our model handles complex multi-room layouts effectively.
• For larger scenes, our approach can be naturally extended by tiling the input into smaller subregions and merging the outputs—an effective strategy in applications such as cadastral mapping and topographic reconstruction. This allows flexible scaling without requiring retraining or modification of the core architecture.

Question 3: Fixed floorplan density map size

Answer:

• We appreciate the suggestion regarding image tiling. The fixed 256×256 resolution is a common design choice in generative models and has proven sufficient for typical residential layouts in datasets like Structured3D and SceneCAD. For larger or more complex spaces, the resolution can be increased or the floorplan can be tiled into smaller subregions—both are practical, widely used strategies that preserve the core architecture.

Question 4: Typos and Mis-links

Answer:

• Thank you for pointing these out. We will correct all typos and fix the broken references in the revised manuscript.

Thank you for the rebuttal and for clarifying that some of the references in the review are already in the paper.

The only limitation of the paper is related to the "amount of contributions" with respect to the closest work "RoomFormer" that are the choice in i) edge representation; ii) network architecture.

The rebuttal highlights how this is contribution from related work but this was already very well written in the paper.

This relatively small limitation is compensated for by the overall quality of it (soundness of the methods, experiments, clarity), so the final rating from reject to borderline accept to represent this final assessment.

Thank you for your valuable comments and suggestions. We will revise the paper accordingly and truly appreciate your recognition and encouragement.

Question 1: Edge-centric representation and post-processing

Answer:

• Thank you for raising this point. While CAGE performs a lightweight edge-to-polygon conversion, it is not an optimization based post-processing step in the traditional sense. The transformer decoder directly predicts directed edge sequences that form closed loops, and the conversion simply connects these edges into polygons deterministically. It does not involve learning, optimization, or iterative refinement.
• Importantly, this conversion is highly robust: in the Structured3D dataset, over 98.71% of intersections are of Types I–III (i.e., directionally valid and unambiguous). The conversion's sensitivity to the threshold (epsilon) is minimal—as shown in the table below, varying epsilon from 0.0001 to 0.2× edge length changes Room metrics by <0.01% and other metrics by less than 0.05%, confirming stability and negligible impact on accuracy. This threshold-based approach offers a practical trade-off between theoretical rigor and engineering feasibility.

Question 2: Ablation studies in supplementary

Answer:

• We appreciate your suggestion and will move the key ablation results from Appendix B.5 to the main paper to clearly highlight the impact of edge representation and denoising decoder.

Question 3: Backbone network comparison

Answer:

• Thanks for your suggestion. We will clarify this in the paper and include each model’s backbone and parameter count in Tables 1 and 2. All our baseline results (for RoomFormer and others) use a ResNet-50 backbone. Our method also uses ResNet-50 in these comparisons. CAGE has 40.96M parameters with a ResNet-50 backbone (comparable to RoomFormer’s 40.60M) and 210M with Swin Transformer V2. Despite this, its lightweight architecture enables fast inference (~0.01s/image), matching RoomFormer and outperforming other baselines in speed.

Question 4: Edge endpoint order

Answer:

• We enforce a consistent edge ordering by starting from the top-left corner of each room polygon, listing edges counter-clockwise, with the first endpoint as the starting point. This sequence is used throughout training to ensure consistency.

Question 5: Redundancy in edge representation

Answer:

• Thank you for the insightful suggestion. We experimented with an explicit co-point regularization loss that encourages adjacent edges to share endpoints. The results (see table below) show modest gains in most metrics, particularly for corner and angle F1 scores, though room recall slightly drops. Without this term, our vertex coordinate loss already implicitly promotes shared vertices by aligning predicted edges to ground truth. We will explore this mechanism behind the interesting result in future work.
• Notably, the explicit regularization improves learning but cannot guarantee exact vertex alignment due to its soft constraint—so edge intersection is still needed during edge-to-polygon conversion.
• We do not consider the edge representation redundant. Each edge encodes position, direction, and length—information not explicitly available in corner-based approaches. This rich geometric encoding enhances global layout reasoning and precision under occlusion.

Question 6: RoomFormer results discrepancy

Answer:

• We reproduced RoomFormer using its official code, hyperparameters, and data splits on our hardware. Slight differences are expected due to implementation environments, but the setup ensures fair ablation comparisons. As shown below, our full model significantly outperforms both the original RoomFormer and our replicated baseline (*), validating the effectiveness of our contributions.

Question 7: Typos and Mis-links

Answer:

• Thank you for pointing these out. We will correct all typos and fix the broken references in the revised version.

Thank you for your valuable comments and suggestions. We will revise the paper accordingly and truly appreciate your recognition and encouragement.

Question 1: PolyRoom better on SceneCAD and cross-dataset generalization precision?

Answer:

• PolyRoom builds on RoomFormer and leverages a segmentation module (Mask2Former) to generate initial query contours, making it a non-end-to-end method. Its strong performance on SceneCAD—a single-room dataset—largely benefits from the segmentation model, which already provides high-quality priors by the segmentation module. However, this reliance limits its ability to encode multiple-room layouts or perform global reasoning, often leading to room-level conflicts. While PolyRoom shows higher precision due to its dependence on segmentation quality, our method achieves superior recall and F1 on key metrics (Room IoU, Corner Recall, Angle Recall, and Angle F1), highlighting stronger generalization and structural robustness.

Question 2: How sensitive is the method to distance threshold selection for Type IV and others?

Answer:

• The edge-to-polygon convertion is empirically stable with respect to the threshold. Type I intersections account for 97.06%, while Types II–IV cover 1.17%, 0.47%, and 1.29%. Shown in the following table, varying the threshold (0.0001 to 0.2× edge length) changes metrics by less than 0.05%, confirming robustness to the epsilon value.

Question 3: Any intuition as to why SwinTransformer degrades in performance when used in isolation?

Answer:

• In our experiments, SwinTransformer captures fine details but is more sensitive to noise, often resulting in irregular or self-intersecting polygons when used directly in the RoomFormer framework. In contrast, our method better leverages Swin’s global reasoning through the edge-based representation and dual-query strategy, enhancing robustness and yielding more stable and accurate layouts.

Question 4: Extremely limited runtime analysis – does CAGE have the same speed as RoomFormer? What does "lightweight architecture and no need for additional refinement" mean? [L261-262]

Answer:

• CAGE has 40.96M parameters with a ResNet-50 backbone (vs. RoomFormer’s 40.60M), converges stably (650 epochs on Structured3D, 400 on SceneCAD), and achieves fast inference (~0.01s/image), matching RoomFormer and outperforming other baselines in speed. We will include a detailed runtime analysis in the revised version.
• The phrase “lightweight architecture and no need for additional refinement” refers to our advantage over all other methods except RoomFormer. CAGE only performs a simple edge-to-polygon conversion, making it faster than models requiring complex refinement or optimization.

Question 5: Since the edge-based representation models room layout as fixed-length sequence, how does this scale with complexity of room layout? And, does this affect performance, both in terms of speed and accuracy?

Answer:

• We use a configuration of 800 tokens (40 polygons × 20 edges), which is sufficient for most residential layouts. For example, Structured3D typically contains fewer than 10 rooms per floor, and SceneCAD includes only one room per case. As demonstrated in case 3410 (Fig. 10, supplementary), our model handles complex multi-room layouts effectively.
• Our method is inherently scalable and built on a general framework. The primary constraint is token length, which is bounded by GPU memory. Scaling is feasible by increasing the token count, though it come with quadratic complexity (O(n²) in both runtime and memory), as is typical for transformer-based models.
• For very larger scenes, our approach can be naturally extended by tiling the input into smaller subregions and merging the outputs—a widely used and effective strategy in applications such as cadastral mapping and topographic reconstruction.

Question 6: Why is there extreme variability in different checkpoints within the same training run? Is this due to convergence related issues of the model?

Answer:

• We observed some variability across checkpoints, likely due to stochastic elements such as dropout. However, these fluctuations are minor, and overall convergence is stable. We repeatedly verified the experiments, and the results reported in the paper are consistent and robust across runs.

Question 7: How does multi-room floor-plan estimation work? Some visual examples would be helpful.

Answer:

• We adopt a two-level token structure where tokens 0–m represent the first polygon, m+1–2m the second, and so on, allowing structured and efficient multi-room representation. We will include additional visual examples to better illustrate this in the revised version.

Question 8: Suggestion: the authors can incorporate training on synthetic datasets such as Aria Synthetic Environments (ASE) to improve generalization precision.

Answer:

• Thank you for the suggestion. ASE’s large-scale synthetic dataset (100K scenes, ~28× larger than Structured3D) is indeed valuable. We are actively collecting diverse real and synthetic scenes, including ASE, to further enhance generalization in future extensions.

Question 9: Typos and Mis-links:

Answer:

• Thank you for pointing these out. We will correct all typos and fix the broken references in the revised version.

Thank you for your valuable comments and suggestions. We will revise the paper accordingly and truly appreciate your recognition and encouragement.

Question 1: Robustness under extreme occlusion

Answer:

• We evaluated CAGE under severe occlusion scenarios, highlighted by red circles in Fig. 6 and Fig. 11 (supplementary). E.g., in Scene 3292, the living room has two fully unscanned walls and one partially visible wall due to the absence of a scanner. In Scene 3356, two walls in the bottom-right room are unscanned, and one wall in the top-right bedroom is fully blocked by a wardrobe. Despite these challenges, CAGE reconstructs complete room polygons, showing strong robustness where corner-based methods often fail.

Question 2: Complex geometries

Answer:

• CAGE is not limited to Manhattan layouts or flat surfaces. As highlighted in red circles in Scene 3493 (Fig. 6) and Scene 3443 (Fig. 10), as well as in many cases in Figs. 7–8, CAGE handles slanted walls by predicting edges at arbitrary angles and approximates curved structures using short segments, similar to other polygon-based methods. While Structured3D and SceneCAD do not contain curved geometry, cross-dataset testing on our collected data confirms CAGE’s generalizability. We will include examples of curved walls in the supplementary material of the revised version.

Question 3: Short wall segments

Answer:

• CAGE captures short wall segments and fine structures effectively, including alcoves and dividers (e.g., the living room in Fig. 6). Its edge-based representation, combined with global reasoning, reduces the risk of missing details that corner-based methods often overlook. Even at 256×256 resolution, CAGE remains robust to noise and occlusion, with potential for further gains at higher resolutions.

Question 4: Computational efficiency

Answer:

• Thank you for the suggestion. We will report full efficiency details in the revised manuscript. CAGE has 40.96M parameters with a ResNet-50 backbone (vs. RoomFormer’s 40.60M), converges stably (650 epochs on Structured3D, 400 on SceneCAD), and achieves fast inference (~0.01s/image), matching RoomFormer and outperforming other baselines in speed.