Hierarchical Solver for Reassembling Mixed Puzzles of Eroded Gaps

Moving beyond the canonical problem of reassembling a single puzzle from all its pieces, this research investigates the simultaneous reconstruction of multiple puzzles. This entails jointly inferring the correct partition of intermingled pieces into their respective sets and the spatial configuration for each set. Inspired by human problem-solving strategies, we propose an Alternating Hierarchical Solver (AHS) comprising a hyper-solver that coarsely groups pieces into semantic clusters, and a local solver that reassembles each cluster via a genetic algorithm. This process is iteratively guided by dynamic visual cues from local pairwise piece correlations and global puzzle reconstruction status. The two modules operate alternately until all puzzles are solved. AHS outperforms state-of-the-art models on two datasets across four settings. Codes are available at https://github.com/sincostar/AHS.