RCLT-Net: An Indoor Panoramic Room Layout Estimation Framework Based on Consecutive Dilated Convolutions and Transformer

Current mainstream models for indoor panoramic room layout depth estimation primarily utilize deep residual networks and various Transformer-based modifications to reconstruct room boundaries from 2D images. However, these methods often encounter challenges such as limited receptive fields, insufficient multi-scale feature interaction and integration capabilities, and blurred boundary estimations in complex scenes. To overcome these challenges, we propose ResCDC-Net, a novel feature extraction model that expands the receptive field by incorporating continuous dilated convolutions within convolutional layers and applying varying dilation rates in residual blocks. This design enhances the perception of fine-grained multi-scale features. Additionally, we introduce the Local-Global Feature Interaction Shifting Window Transformer, which significantly improves the Transformer model’s ability to extract features by fostering robust interactions between local and global features, enabling more flexible and efficient feature representation. Furthermore, we employ a triplet loss function based on positive and negative anchor points to ensure clear boundary distinctions in complex object boundary regions. Extensive experiments validate that our proposed RCLT-Net surpasses SoTA models in depth estimation accuracy and demonstrates exceptional performance across multiple challenging public datasets.