Deep Learning of Graph Matching 閱讀筆記
阿新 • • 發佈：2018-12-31
We have presented an end-to-end learning framework for graph matching with general applicability to models containing deep feature extraction hierarchies and combinatorial optimization layers. We formulate the problem as a quadratic assignment under unary and pair-wise node relations represented using deep parametric feature hierarchies. All model parameters are trainable and the graph matching optimization is included within the learning formulation. As such, the main challenges are the calculation of backpropagated derivatives through complex matrix layers and the implementation of the entire framework (factorization of the affinity matrix, bi-stochastic layers) in a computationally efficient manner. Our experiments and ablation studies on diverse datasets like PASCAL VOC keypoints, Sintel and CUB show that fully learned graph matching models surpass nearest neighbor counterparts, or approaches that use deep feature hierarchies that were not refined jointly with (and constrained by) the quadratic assignment problem.