ArchCAD-400K: A Large-Scale CAD drawings Dataset and New Baseline for Panoptic Symbol Spotting

Q1: About the release of the dataset and code.

A1: We fully intend to release all major components to the community.

More specifically:

(1) The dataset will be publicly released on HuggingFace before October 2025. We are currently undergoing final legal and technical checks to ensure proper anonymization and licensing.

(2) In addition to the data, we will also release a toolkit designed for parsing, cleaning, and managing both our dataset and other CAD drawings from diverse sources. We believe this will significantly facilitate downstream research and application.

(3) The code for the proposed DPSS framework will also be made available on GitHub along with the dataset toolkit.

Q2: About the reporting of training time, computational resources used, and inference time.

A2: Thank you for the valuable suggestion. We agree that reporting computational efficiency is important for a comprehensive understanding of the proposed method.

We have included a summary of the training time and inference speed in the revised version as follows. Specifically, all experiments were conducted using 8 NVIDIA A800 GPUs. For the ArchCAD-400K dataset, we employed a batch size of 4 per GPU and trained the model for 10 epochs. For the FloorPlanCAD dataset, we used a batch size of 2 per GPU with 50 training epochs. Further details of the training and inference setup can be found in Section 6.1 of the main text.

Q3: About the framework’s performance on noisy or real-world CAD environments beyond curated datasets.

A3: We appreciate the reviewer's concern regarding data realism and robustness.

First, regarding layer information, during the design of the panoptic symbol spotting task, we explicitly accounted for the possibility of incomplete or inconsistent layer annotations. Accordingly, our algorithmic evaluations were conducted under both with-layer and no-layer settings, as reflected in the "Additional prior inputs" column of Tables 2 and 3 in the main paper.

Second, with respect to noise and imperfections in the drawings, our ArchCAD-400K dataset is collected directly from real-world architectural design archives, and thus naturally reflects the challenges and noise present in practical CAD environments.

In summary, we have made a deliberate effort to ensure that both our dataset and experimental setup reflect realistic conditions encountered in actual CAD workflows.

Q4: About reporting the standard deviations and multiple experimental runs.

A4: We acknowledge this limitation. Due to the considerable computational cost associated with training (as noted in Q2), it is difficult to perform multiple full runs for each setting. We plan to include repeated experiments for key configurations in the revised version to further validate the stability of our results.

Nevertheless, to partially mitigate this, we used consistent data splits and evaluation protocols across all methods to ensure fair comparison. Besides, the performance gaps between our method and baselines are obvious across both datasets (Table 2, 3, 4), which we believe provides sufficient support for the reliability of our claims.

References

[1] Zhiwen Fan, Tianlong Chen, Peihao Wang, and Zhangyang Wang. Cadtransformer: Panoptic symbol spotting transformer for cad drawings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10986–10996, 2022.

[2] Wenlong Liu, Tianyu Yang, Yuhan Wang, Qizhi Yu, and Lei Zhang. Symbol as points: Panoptic symbol spotting via point-based representation. arXiv preprint arXiv:2401.10556, 2024.

[3] Wenlong Liu, Tianyu Yang, Qizhi Yu, and Lei Zhang. Sympoint revolutionized: Boosting panoptic symbol spotting with layer feature enhancement. arXiv preprint arXiv:2407.01928, 2024.

Q1: About the reduction in categorization granularity compared to previous datasets, such as merging multiple categories into "furniture".

A1: Thank you for the observation. We intentionally chose to reduce the categorization granularity for certain object types. For example, we grouped subtypes such as tables, chairs, and cabinets into a single furniture category. Our rationale is twofold:

(1) Limited Utility: While structural elements such as walls, columns, and curtain walls are essential for architectural modeling and reconstruction, fine-grained distinctions among furniture types offer relatively limited benefit for these purposes. The impact of individual furniture subtypes on the global structure is minimal, making detailed categorization a lower priority.

(2) Annotation Efficiency: Preliminary analysis revealed that the furniture class could be subdivided into over 20 distinct subtypes. However, annotating such fine-grained labels across large-scale datasets would introduce substantial overhead. Given the limited structural value of these distinctions, we opted for a more efficient, coarser labeling strategy.

Overall, this reflects a conscious trade-off between annotation cost and semantic value, aligned with the objectives of our reconstruction task.

Q2: About the relationship between dataset size, the number of distinct drawings, and potential diversity bias due to the inclusion of larger buildings.

A2: We appreciate this important point. Our dataset contains a larger number of distinct drawings than FloorPlanCAD. FloorPlanCAD has 1,332 distinct CAD drawings. In contrast, our dataset includes 5,538 distinct drawings.

Both datasets are organized in a chunk-wise fashion, but the increased number of large buildings in our dataset reflects an expansion in both spatial coverage and complexity, rather than a reduction in diversity. We acknowledge the concern that larger buildings could skew the distribution; however, we take steps to mitigate this, as described in our response to Q3.

Q3: About the basis of distribution in Figure 4(a), and the potential bias introduced by chunk-wise representation.

A3: Thank you for raising this concern. Figure 4(a) indeed shows the distribution by source drawings, not by chunks. We recognize that the correlation between drawing type and spatial size could result in an overrepresentation of certain sources at the chunk level.

To address this, we conducted an additional analysis of the dataset with a chunk-wise breakdown, presented in the as follows.

This analysis revealed that while there are some differences in proportions between drawing-wise and chunk-wise distributions, the semantic category distribution and spatial diversity remain well-balanced at the chunk level.

Q4: About the generation of initial annotation proposals.

A4: We appreciate the interest in our annotation workflow. The initial labeling proposals were generated entirely through CAD layer-based heuristics, without any machine learning or predictive components.

Specifically, we built a layer-to-category mapping table by consolidating naming conventions across various CAD standards and companies. This mapping enabled automatic pre-labeling of most graphical elements. Annotators then reviewed and refined these proposals based on context and visual correctness.

A partial snapshot of this mapping table is provided in Appendix A.2. This rule-based pipeline ensured high initial labeling quality and interpretability, which was subsequently verified through multi-stage quality control.

Q5: About the qualitative results.

A5: We appreciate the reviewer's suggestion. In the revised version, we will include additional qualitative visualizations to better illustrate the strengths and limitations of our method. These visual examples will be carefully selected to correspond with the quantitative metrics, thereby providing more intuitive insights into what the reported numerical values reflect in practice.

Q6: About the definition and usage of the terms "Thing" and "Stuff" in evaluation metrics.

A6: Thank you for highlighting this point. The terms “Thing” and “Stuff” come from definition of panoptic symbol spotting [1].

"Things" refer to countable, instance-level objects (e.g., columns, doors).

"Stuff" refers to amorphous symbols of similar objects or material with semantic meaning (e.g., wall).

Evaluating the two metrics separately allows for a more precise assessment of the model's performance on semantic and instance segmentation tasks, which is a common practice on this problem [1–3].

Q7: About the handling of intellectual property in the dataset.

A7: We acknowledge your concern and agree that raw floorplans could constitute IP, even if scrubbed of metadata and labels. During the data processing period, we applied several measures for IP protection:

(1) Spatial Chunking: Floorplans are divided into localized chunks, containing only partial spatial context.

(2) Cross-Source Shuffling: Chunks are decoupled from their original drawing context and mixed across different sources.

(3) Text Anonymization: Any textual elements that could reveal project background (like project names, client identifiers, etc.) are removed.

Due to this process, it is extremely difficult to reconstruct the original full floorplan, even with access to all chunks. Some of the shuffled chunks are shown in supplementary materials. We believe this approach significantly reduces the risk of re-identification or IP leakage.

References

[1] Zhiwen Fan, Lingjie Zhu, Honghua Li, Xiaohao Chen, Siyu Zhu, and Ping Tan. Floorplancad: A large-scale cad drawing dataset for panoptic symbol spotting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10128–10137, 2021.

[2] Zhiwen Fan, Tianlong Chen, Peihao Wang, and Zhangyang Wang. Cadtransformer: Panoptic symbol spotting transformer for cad drawings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10986–10996, 2022.

[3] Wenlong Liu, Tianyu Yang, Yuhan Wang, Qizhi Yu, and Lei Zhang. Symbol as points: Panoptic symbol spotting via point-based representation. arXiv preprint arXiv:2401.10556, 2024.

Q1: About computational speed of the proposed method.

A1: We appreciate the reviewer's suggestion regarding the comparison of computational speed with other SOTA methods.

Our model operates on images with a resolution of 700×700 pixels, which introduces some additional computational cost, but the increase is modest due to the relatively small size. We selected this resolution empirically—higher resolutions did not yield further performance gains, while lower resolutions led to noticeable performance degradation.

We also found that incorporating image features significantly improves model performance. These features enhance the robustness to noisy primitives and contribute positively to overall panoptic quality, making the moderate computational overhead worthwhile.

We measured the average inference latency of several representative methods on an NVIDIA A800 GPU, as summarized in the table below. Compared to SymPoint-V2[3], our method increases latency by only 30%, while achieving a 3.0% and 10.1% improvement in panoptic quality on ArchCAD-400K and FloorPlanCAD respectively.

Q2: About the impact of loss function weighting.

A2: We conduct some experiments to determine a reasonable configuration of loss function. Specifically, we experimented with different combinations of loss weights $\lambda_{cls}$ : $\lambda_{bce}$ : $\lambda_{dice}$, and empirically identified a reasonable range for balancing these terms. The setting around 1:2:2 yielded stable performance and was adopted as the default in our implementation.

Q3: About the release of the dataset.

A3: We fully intend to release all major components to the community.

More specifically:

(1) The dataset will be publicly released on HuggingFace before October 2025. We are currently undergoing final legal and technical checks to ensure proper anonymization and licensing.

(2) In addition to the data, we will also release a toolkit designed for parsing, cleaning, and managing both our dataset and other CAD drawings from diverse sources. We believe this will significantly facilitate downstream research and application.

(3) The code for the proposed DPSS framework will also be made available on GitHub along with the dataset toolkit.

References

[1] Zhiwen Fan, Tianlong Chen, Peihao Wang, and Zhangyang Wang. Cadtransformer: Panoptic symbol spotting transformer for cad drawings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10986–10996, 2022.

[2] Wenlong Liu, Tianyu Yang, Yuhan Wang, Qizhi Yu, and Lei Zhang. Symbol as points: Panoptic symbol spotting via point-based representation. arXiv preprint arXiv:2401.10556, 2024.

[3] Wenlong Liu, Tianyu Yang, Qizhi Yu, and Lei Zhang. Sympoint revolutionized: Boosting panoptic symbol spotting with layer feature enhancement. arXiv preprint arXiv:2407.01928, 2024.

Q1: About the acquisition details of the dataset.

A1: Thank you for pointing this out. The raw CAD data was collected from engineering drawings provided by top-tier architectural design institutions. This ensures both the professional quality and diversity of the dataset, making it representative of real-world CAD design practices.

The data was made available to us under a collaborative research agreement and no direct monetary cost was involved in acquiring the dataset.

Q2: About the concept of layer and the method used for layer extraction.

A2: In CAD drawings, a layer refers to a categorical division of graphical elements defined by engineers. The purpose of layer differentiation is to facilitate clearer organization of drawing content and enable selective rendering of specific components, thereby improving clarity and supporting more effective collaboration among team members.

In our pipeline, we extract layer information by programmatically parsing the CAD files, leveraging metadata embedded in the DWG/DXF formats. These layers are then used to isolate structural components for downstream analysis.

Q3: About the definition and clarification of the term "standardized" drawings.

A3: We appreciate the reviewer's careful reading. The term "standardized" drawings refers to CAD drawings that follow a consistent layering convention, where each type of building component is assigned to a specific layer based on an internally defined standard. This standard ensures semantic consistency across drawings and facilitates automated parsing. A brief explanation of this standard is provided in Appendix A.2. We will refine this statement in the revised version.

Q4: About the reliability of annotations, including inter-annotator agreement and gold standard validation.

A4: We employed several strategies to ensure the reliability and accuracy of the annotations during the annotation process:

(1) Annotation Guidelines: A comprehensive annotation manual was developed, detailing the definition and visual examples of each category. Annotators were trained using this guide to maintain consistency. A portion of the manual is included in Appendix A.1.

(2) Quality Control: All annotations underwent a secondary verification step. Given the complexity of CAD drawings, we implemented a hierarchical review process, where a senior annotator re-checked each annotated file. This procedure helped to eliminate labeling inconsistencies and reduce human errors.

(3) Customized Annotation Pipeline: As mentioned in paper 3.3, before manual labeling, we apply a pipeline that extracts inherent information such as layers and blocks to generate annotation suggestions. This avoids labeling from scratch, improves efficiency, and reduces human errors, especially in densely labeled CAD drawings.

Q5: About the semantic categories.

A5: The semantic categories used in our project were independently defined based on domain requirements. Our primary criterion was whether the categorized elements are essential for reconstructing a Building Information Model (BIM). While some categories align with prior work such as FloorPlanCAD, our classification scheme was tailored to industrial CAD drawings and encompasses additional component types specific to structural and architectural design.

Q6: About the train/val/test split of the dataset.

A6: Thank you for raising this important point. We split the dataset at the chunk level, rather than the drawing level, to form the train/val/test sets. This decision was made based on two key considerations:

(1) Low risk of data leakage: Chunks assigned to different splits do not spatially overlap, which theoretically prevents any form of data leakage between training and testing. Although some chunks may come from the same large drawing, their sparse details result in insignificant correlation between chunks, keeping the risk of data leakage low.

(2) Improved balance across splits: Variation in both class distribution and drawing area can cause imbalance when splitting at the drawing level. Chunk-level splitting helps ensure better balance in instance counts and spatial coverage across splits.

As a result, class balance is easier to achieve and more controllable under the chunk-level splitting strategy.

Q7: About the relationship between symbol complexity and model performance.

A7: Thank you for raising this insightful point. We agree that understanding how symbol complexity and image resolution affect model performance is an interesting and helpful for understanding the problem.

To investigate the impact of symbol complexity, we analyzed the number of primitives within each symbol as a proxy for complexity. Specifically, we partitioned the test symbols into eight subsets based on the percentile of primitive count and then computed evaluation metrics on each subset.

From the results, we can find that there exists a certain correlation between symbolic complexity and model performance. In brief, both extremely simple and extremely complex symbols represent performance bottlenecks for the model.

For symbols with low complexity within the range [1, 4), the model exhibits relatively low recognition quality (RQ), likely due to the lack of distinctive structural features. Within the moderate complexity range [4, 46), the model demonstrates relatively stable performance. However, in the highest complexity split [46, ), the performance degradation is primarily attributed to a decline in segmentation quality (SQ), as the high structural complexity leads to intricate and ambiguous symbol boundaries in the symbolic space.

Q8: About the relationship between image resolution and model performance.

A8: Thank you for raising this insightful point. We conducted a set of experiments to evaluate the influence of raster image resolution on model performance. Specifically, we varied the resolution used for rasterizing the input vector graphics while keeping all other components of the pipeline unchanged. The results are summarized as follows.

The results suggest that increasing the resolution is beneficial to model performance when starting from a relatively low baseline. However, beyond a certain threshold further increasing the resolution yields diminishing returns and may lead to degradation in efficiency.

This phenomenon can be attributed to the inherent characteristics of CAD drawings: since they consist primarily of discrete, line-based primitives, the amount of spatial detail is inherently limited. Once the resolution is sufficient to capture and distinguish these primitives, further increases do not yield additional useful information.

Q9: The evaluation metrics are defined in pixel space or symbolic space?

A9: Thank you for raising this question. The evaluation metrics for panoptic symbol spotting are defined in the symbolic space. Their formal definitions can be found in [1], and they are commonly adopted as standard practice in prior work in the panoptic symbol spotting task. More specifically, the key distinction between evaluation in pixel space and symbolic space lies in the way IoU is computed. For a symbol $s$ represented as a set of primitives $s= \lbrace e_1,e_2,… \rbrace$, the IoU between a predicted symbol $s_p$ and a ground truth symbol $s_g$ is calculated as:

$$\mathrm{IoU}(s_p, s_g) = \frac{\sum_{e_i \in s_p \cap s_g} \log{(1+L(e_i))}}{\sum_{e_i \in s_p \cup s_g} \log{(1+L(e_i))}}$$

where $L(e_i)$ denotes the length of the primitive $e_i$.

This formulation ensures that the metric is sensitive to geometric correctness at the symbol level, while also addressing the issue of symbol size imbalance. The use of a logarithmic scaling function $\log{(1+L(e_i))}$ prevents large symbols from disproportionately dominating the metric.

References

[1] Zhiwen Fan, Lingjie Zhu, Honghua Li, Xiaohao Chen, Siyu Zhu, and Ping Tan. Floorplancad: A large-scale cad drawing dataset for panoptic symbol spotting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10128–10137, 2021.