For visual odometry in low-texture environments, dense and semi-dense matching methods are able to provide more reliable correspondences compared to detector-based methods, which typically entails the detection and matching of keypoints. However, detector-free methods such as LoFTR and ASpanFormer often treat the input image-pair unequally within their coarse-to-fine matching strategy. The inter-frame inconsistency caused by fixing feature locations in one image and seeking sub-pixel dense matches in another, hinders its use in downstream optimization for tasks like pose estimation, ultimately restricting gains in localization accuracy. To mitigate this problem, we improve it from two perspectives. The first is a decentralized midpoint cross-search strategy in the matching stage, which treats each input image equivalently. The second involves a novel loss function designed by extending the input image-pair to multiple consecutive frames and incorporating their consistency constraints. Furthermore, this paper presents a Transformer-based pose estimation network, which utilizes raw image content information to enhance dense matches. The effectiveness of our method, particularly in terms of inter-frame consistency and localization accuracy, is demonstrated through experiments on both public and self-collected datasets. The code is available at https://github.com/2025Night/DFTransVO .