Efficient interactive segmentation of three-dimensional Gaussians with optimal view selection

Three-dimensional (3D) scene representation has advanced rapidly in recent years, drawing the focus of more researchers. One of the main challenges for researchers is quickly and accurately segmenting 3D objects. Previous work has achieved excellent segmentation accuracy, but retraining requires a significant amount of time. Additionally, most methods fail to provide users with an efficient and convenient segmentation experience. To address these issues, we present Efficient Interactive Segmentation of 3D Gaussians (EISG), an efficient interactive segmentation method that eliminates the need for lengthy retraining. We first design an optimal view selection (OVS) method. This method uses 3D Gaussian entropy and image uncertainty to evaluate the quantity of view information. OVS helps users quickly select the optimal segmentation view, thereby enhancing interaction efficiency. Secondly, we use projection to find the target foreground rapidly and then segment the approximate objects using a clustering algorithm. Thirdly, we design a spatial-color background filter (SCBF) using the depth and color of 3D Gaussians. SCBF enables precise segmentation without needing retraining. Our method has been systematically tested on multiple datasets. Compared to other methods, the results demonstrate that EISG achieves ideal accuracy while significantly reducing processing time.