Joint Multiple Image Parametric Transformation Estimation Via Convolutional Neural Networks

The correspondence problem is conventionally performed at the pairwise level, i.e., finding the correspondence model, e.g., affine transformation between two input images. While, this paper tackles the scenario when more than two images, e.g., a sequence of images are considered either for model learning or inference. Our proposed approach is based on the recent work on convolutional neural network for geometric matching model. Specifically, we extend this baseline by introducing sequential cycle consistency check that can involve multiple images. The learning is performed in a supervised setting provided with ground truth parametric transformation information, while it meanwhile leverages the consistency information as a regularizer during learning. Extensive experiments are performed on the public benchmark dataset, whereby qualitative and quantitative results are both presented. Our method improves the two-image geometric matching network learning baseline by fusing more than two images' information during learning, while it can still be applied for two-image matching for testing.