Accurately recovering 3D shapes from single images that contain complex backgrounds remains a longstanding and difficult task. Unlike machines, the human visual system can effortlessly filter out background distractions and utilize extensive geometric and semantic knowledge to interpret 3D structures precisely. In contrast, current single-image 3D reconstruction methods often struggle to focus on the target object when confronted with complex backgrounds, as noise and irrelevant objects reduce reconstruction accuracy. To address this issue, we propose a cross-modal fusion strategy that integrates feature enhancement with difference-guided attention to enable high-quality 3D reconstruction from single complex images (named DGGR-Net). Specifically, we utilize retrieved 3D models from the ShapeNet dataset as structural priors and introduce a local geometry-preserving graph convolution module (LGPConv) to optimize fine-grained point cloud structures. Additionally, we design a bidirectional spatial attention (BSA) module to effectively capture spatial image features, reducing background interference during feature extraction. Furthermore, we propose a difference-guided cross-modal attention (DCA) module, which explicitly computes the differences between image and point cloud features to guide precise cross-modal feature fusion, thereby improving modality complementarity and robustness. The experimental results show that our proposed method achieves a Chamfer Distance (CD) of \(3.18 \times 10^{-2}\) and an Earth Mover’s Distance (EMD) of \(3.70 \times 10^{-2}\) on the ShapeNet dataset, and \(5.62 \times 10^{-2}\) and \(7.30 \times 10^{-2}\) respectively on the Pix3D dataset. Compared with the latest method RGB2Point, our method achieved an average improvement of approximately 13.32% and 12.55% in both CD and EMD metrics across the ShapeNet and Pix3D datasets.

Augmented reality (AR) has become a research focus in computer vision and graphics, with growing applications driven by advances in artificial intelligence and the emergence of the metaverse. Panoramic cameras offer new opportunities for AR due to their wide field of view but also pose significant challenges for camera pose estimation because of severe distortion and complex scene textures. To address these issues, this paper proposes a lightweight, unsupervised deep learning model for panoramic camera pose estimation. The model consists of a depth estimation sub-network and a pose estimation sub-network, both optimized for efficiency using network compression, multi-scale rectangular convolutions, and dilated convolutions. A learnable occlusion mask is incorporated into the pose network to mitigate errors caused by complex relative motion. Furthermore, a panoramic view reconstruction model is constructed to obtain effective supervisory signals from the predicted depth, pose information, and corresponding panoramic images and is trained using a designed spherical photometric consistency loss. The experimental results demonstrate that the proposed method achieves competitive accuracy while maintaining high computational efficiency, making it well-suited for real-time AR applications with panoramic input.