Abstracting human-made 3D models by a set of primitives, such as cuboid abstraction, is a fundamental task in 3D shape modelling and analysis. Traditionally, different forms of representations, such as edges, volumes, or curves, were used as primitives. Although methods that apply local operations to compute such primitives can produce satisfactory results with their own merits, the computations can be very slow with complex models. Learning-based abstraction methods are much faster but cannot guarantee the fitting precision between the primitives and the original shape. To solve this problem, we propose an unsupervised learning approach for shape abstraction. Our method’s key idea is to use an iterative error feedback (IEF)-based network to improve primitive precision. Our method contains two main steps. First, we use a regression network to predict the initial primitives. Second, we increase the accuracy of the initial primitives by using an IEF-based network, which iteratively outputs the primitive updates. We demonstrate the advantages of our method by comparing it to existing state-of-the-art methods. We also thoroughly evaluate our method by ablation studies.