Numerous multi-modality approaches employ depth completion to address the challenge of imprecise feature fusion between point clouds and images, chiefly attributed to the sparsity of point clouds. However, a notable drawback of this approach is the introduction of excessive redundant operations and noise data throughout the completion process. In this paper, we present a method called Object Detection with Partial Depth Completion (ODPDC), which leverages the inherent knowledge of images to carry out targeted depth completion for detected objects within the images, thereby efficiently minimizing unnecessary operations and mitigating noise data. Our method initially employs a segmentation module to isolate objects detected in the image, guiding the point cloud toward partial depth completion and reducing redundant points and noise. Additionally, we develop a feature extractor named Pseudo Point Convolution (PPC), which adeptly extracts 2D and 3D features from the pseudo point cloud resulting from partial depth completion. In addition, we introduce a novel 3D feature fusion model named Multi- layer Attentional Feature Fusion (MAFF), which effectively facilitates precise feature fusion both locally and globally. Our experimental assessment on the KITTI dataset [1] validates that our proposed approach effectively addresses the challenge of incorporating an excessive number of redundant and noisy data points. Furthermore, the feature extraction and feature fusion modules significantly boost the detection performance.