Various image-based building object recognition approaches have been developed to create as-is Building Information Models (BIMs) of existing buildings. However, existing approaches generally rely on human-designed features to automatically or semi-automatically recognize building objects, which makes them sensitive to input images and difficult to extend to new building objects. Furthermore, when constructing object geometries, most of these approaches are limited to rectangular or pre-defined surface shapes. To address these limitations, this study presents a human-designed features-free, shape constraints-free and fully automatic approach to construct as-is BIM objects from images of a building. This approach adapts Mask R-CNN, a deep convolutional neural network, to automatically recognize and segment building objects with arbitrary shapes (i.e., surface boundary shapes) from images. The segmented objects, characterized by object types and pixel masks, are further geometrically fitted to construct surface geometries. Finally, the constructed building objects are defined in the Industry Foundation Classes (IFC) data format. Three types of building objects (i.e., walls, doors, and lifts) are used in this study. Total 430 images containing these objects collected from the interiors of 4 university buildings are used to train and test the Mask R-CNN model. The test results show that the trained model is accurate and robust to recognize and segment all the three types of building objects. Furthermore, the feasibility of the proposed approach is preliminarily validated by successfully extracting IFC building objects from an image.