Fractured object reassembly is a challenging problem in computer vision with broad applications in industrial manufacturing, archaeology, etc.. Traditional procedural methods rely on local shape descriptors or geometric registration, which are not always robust given the small fraction of fracture faces among fragments. While recent deep learning based methods have shown promising results by incorporating semantic information, they often assume that input fragments are aligned in a canonical pose. In this paper, we propose an approach that eliminates this implicit assumption by predicting shape reassembly results under arbitrary poses. Instead of directly regressing the canonical fragment poses, our neural network predicts the complementary shape of one input fragment given the other fragment to expand potential overlapping areas for later registration.