Abstract: Due to the diversity and density of object types in indoor scenes, as well as issues such as occlusion leading to data loss, the challenge of environmental perception and reconstruction for computers has increased. Focusing on complex indoor point cloud scenes, we propose an method for indoor scene reconstruction driven by object detection. First, we build upon VoteNet, designed a fusion convolutional pooling module to extract richer local features and designing a voting weight module based on an attention mechanism to enhance the attention to foreground features. Additionally, we utilize an object relationship module to learn spatial relationship features, optimizing indoor object detection results. We then extract effective object point clouds based on the Intersection over Union of bounding boxes and employ optimal model retrieval algorithms to obtain more accurate matching models from the model repository. Finally, using the instance information of the objects and the retrieved model information as input, we optimize the model pose through geometric spatial constraints of the surrounding environment, resulting in more reasonable and refined indoor scene reconstruction outcomes. Experimental results on the ScanNet dataset show that the proposed method achieves an average precision of 63.8%, an improvement of 7.0 percentage points over VoteNet. This method demonstrates robust performance and accuracy, particularly in addressing issues related to low matching precision caused by object incompleteness.