Although convolution neural network based stereo matching architectures have\nmade impressive achievements, there are still some limitations: 1)\nConvolutional Feature (CF) tends to capture appearance information, which is\ninadequate for accurate matching. 2) Due to the static filters, current\nconvolution based disparity refinement modules often produce over-smooth\nresults. In this paper, we present two schemes to address these issues, where\nsome traditional wisdoms are integrated. Firstly, we introduce a pairwise\nfeature for deep stereo matching networks, named LSP (Local Similarity\nPattern). Through explicitly revealing the neighbor relationships, LSP contains\nrich structural information, which can be leveraged to aid CF for more\ndiscriminative feature description. Secondly, we design a dynamic\nself-reassembling refinement strategy and apply it to the cost distribution and\nthe disparity map respectively. The former could be equipped with the unimodal\ndistribution constraint to alleviate the over-smoothing problem, and the latter\nis more practical. The effectiveness of the proposed methods is demonstrated\nvia incorporating them into two well-known basic architectures, GwcNet and\nGANet-deep. Experimental results on the SceneFlow and KITTI benchmarks show\nthat our modules significantly improve the performance of the model.\n