**CAD-SIGNet: CAD Language Inference from Point Clouds using Layer-wise Sketch Instance Guided Attention** **Authors:** Mohammad Sadil Khan, Elona Dupont, Sk Aziz Ali, Kseniya Cherenkova, Anis Kacem, Djamila Aouada **Abstract:** Reverse engineering in Computer-Aided Design (CAD) aims to uncover the CAD process behind a physical object given its 3D scan. This paper proposes CAD-SIGNet, an end-to-end trainable and auto-regressive architecture to recover the design history of a CAD model represented as a sequence of sketch-and-extrusion steps from an input point cloud. The model learns CAD visual-language representations through layer-wise cross-attention between the point cloud and CAD language embedding. A new Sketch Instance Guided Attention (SGA) module is introduced to reconstruct fine-grained sketch details. CAD-SIGNet not only reconstructs the full design history but also provides multiple plausible design choices, enabling interactive reverse engineering. Extensive experiments on public CAD datasets demonstrate the effectiveness of CAD-SIGNet in full design history recovery and conditional auto-completion from point clouds. **Contributions:** - An end-to-end trainable auto-regressive network for inferring CAD language from point clouds. - Multi-modal transformer blocks with layer-wise cross-attention between point cloud and CAD language embedding. - A SGA module for guiding cross-attention to relevant regions of the point cloud for sketch parameterization. - Thorough experimental validation in two reverse engineering settings. **Related Work:** - Deep learning-based CAD reverse engineering focuses on recovering geometric features from point clouds. - CAD as a language is represented using language modeling techniques, with some works focusing on language modeling of CAD sketches and others on modeling CAD models. - Previous works like DeepCAD and MultiCAD infer CAD language using feed-forward strategies or separate modality learning, lacking the ability to provide multiple design choices. **CAD-SIGNet Architecture:** - The architecture consists of multi-modal transformer blocks with layer-wise cross-attention. - The SGA module enhances sketch parameterization by focusing attention on specific points within a sketch instance. - Training and inference strategies are described, including hybrid sampling for generating multiple plausible design sequences. **Experimental Results:** - CAD-SIGNet outperforms existing methods in design history recovery and conditional auto-completion tasks. - Ablation studies show the importance of the SGA module and hybrid sampling. - Real-world applications, including cross-dataset experiments and realistic 3D scans, demonstrate the model's effectiveness. **Conclusion:** CAD-SIGNet is an effective tool for recovering CAD design history from point clouds, offering multiple design choices and improving the accuracy of 3D reconstructions.**CAD-SIGNet: CAD Language Inference from Point Clouds using Layer-wise Sketch Instance Guided Attention** **Authors:** Mohammad Sadil Khan, Elona Dupont, Sk Aziz Ali, Kseniya Cherenkova, Anis Kacem, Djamila Aouada **Abstract:** Reverse engineering in Computer-Aided Design (CAD) aims to uncover the CAD process behind a physical object given its 3D scan. This paper proposes CAD-SIGNet, an end-to-end trainable and auto-regressive architecture to recover the design history of a CAD model represented as a sequence of sketch-and-extrusion steps from an input point cloud. The model learns CAD visual-language representations through layer-wise cross-attention between the point cloud and CAD language embedding. A new Sketch Instance Guided Attention (SGA) module is introduced to reconstruct fine-grained sketch details. CAD-SIGNet not only reconstructs the full design history but also provides multiple plausible design choices, enabling interactive reverse engineering. Extensive experiments on public CAD datasets demonstrate the effectiveness of CAD-SIGNet in full design history recovery and conditional auto-completion from point clouds. **Contributions:** - An end-to-end trainable auto-regressive network for inferring CAD language from point clouds. - Multi-modal transformer blocks with layer-wise cross-attention between point cloud and CAD language embedding. - A SGA module for guiding cross-attention to relevant regions of the point cloud for sketch parameterization. - Thorough experimental validation in two reverse engineering settings. **Related Work:** - Deep learning-based CAD reverse engineering focuses on recovering geometric features from point clouds. - CAD as a language is represented using language modeling techniques, with some works focusing on language modeling of CAD sketches and others on modeling CAD models. - Previous works like DeepCAD and MultiCAD infer CAD language using feed-forward strategies or separate modality learning, lacking the ability to provide multiple design choices. **CAD-SIGNet Architecture:** - The architecture consists of multi-modal transformer blocks with layer-wise cross-attention. - The SGA module enhances sketch parameterization by focusing attention on specific points within a sketch instance. - Training and inference strategies are described, including hybrid sampling for generating multiple plausible design sequences. **Experimental Results:** - CAD-SIGNet outperforms existing methods in design history recovery and conditional auto-completion tasks. - Ablation studies show the importance of the SGA module and hybrid sampling. - Real-world applications, including cross-dataset experiments and realistic 3D scans, demonstrate the model's effectiveness. **Conclusion:** CAD-SIGNet is an effective tool for recovering CAD design history from point clouds, offering multiple design choices and improving the accuracy of 3D reconstructions.