Object Detection-Driven 3D Indoor Scene Reconstruction

Key Points

• The aim is to improve 3D indoor scene reconstruction accuracy by leveraging object detection techniques.
• Utilized VoteNet as a base for enhanced local feature extraction using a convolution pooling module.
• Designed a voting weight module to focus on foreground features and learn spatial relationship features.
• Optimized model pose based on geometric constraints of the environment using instance and retrieval information.
• Achieved an average accuracy of 63.8% on the ScanNet dataset, an improvement of 7.0 percentage points over VoteNet.
• Demonstrated robustness and accuracy, especially in handling low-matching precision due to object occlusions.