Conference Agenda

This work introduces RE-CAD, a pipeline that converts multiview images into editable Computer-Aided Design (CAD) command sequences. Building on ARE-Net and DeepCAD, it addresses the lack of recoverable CAD design procedures and the scarcity of paired image–CAD sequence data. A synthetic multi-view dataset was therefore generated from DeepCAD models using an automated Blender pipeline, with a black background and a High-Dynamic-Range imaging (HDRI) variant; the latter adds ray tracing, reflections, and depth blur to approximate real-world imaging conditions.

The architecture uses a pretrained ResNet-18 encoder to process 12 views per object, a Gated Recurrent Unit (GRU) pooler to aggregate their features, and the pretrained DeepCAD decoder. The encoder–pooler was trained by Mean Squared Error (MSE) against 256-dimensional target vectors produced by the pretrained DeepCAD encoder. Hyperparameters were optimized through two Optuna studies, followed by multiple training sessions and 5-fold cross-validation. The reported losses evaluate latent-space alignment, not command validity or geometric reconstruction accuracy. Validation loss plateaued while training loss decreased; an exploratory comparison also showed a lower, more stable trend in validation loss for the 8% subset than for the 5% subset.