Research on 3D Modeling Based on Improved YOLOv7—A Case Study of Nanshitou Water Supply Department Site in Guangzhou

In view of the problems of different scales, missing surfaces and irrecoverable photos of architectural sites, this paper puts forward an improved algorithm based on YOLOv7 (YOLOv7-ESL). This algorithm combines the attention mechanism of ECA channel and Swin Transformer V2 module, and makes full use of the edge contour of the building, and designs the fusion of the edge feature map and the original network feature map, which significantly improves the detection ability of building sites and the efficiency of detail feature extraction. Experimental results show that YOLOv7-ESL is better than the original YOLOv7 model in terms of detection accuracy (94.8% for mAP@0.5) and real-time performance (FPS of 29.8 frames/s). In the process of SketchUp 3D modeling, by applying the improved YOLOv7 algorithm to the image analysis of the site, the accurate identification and positioning of various building structures, detailed features and potential damage areas in the site were successfully realized, which can not only show the macro layout of the site in detail, but also accurately reproduce the micro details. It provides a more advanced and practical digital path for cultural relics protection, restoration planning and heritage display.