Total: 1
Establishing dense correspondences is crucial yet computationally demanding in multi-view tasks. Although coarse-to-fine schemes mitigate computational costs, their efficiency remains limited by the substantial demands of heavy feature extractors and global matchers. In this paper, we propose Adaptive Refinement Gathering, a refinement pipeline that reduces reliance on these costly components without sacrificing accuracy. The pipeline consists of (i) a content-aware offset estimator that leverages content information for lightweight correlation volume encoding and decoding; (ii) a locally consistent match rectifier robust to large global initial errors; (iii) a locally consistent upsampler that yields fewer artifacts around depth-discontinuous edges. Additionally, we introduce an adaptive gating strategy that, in conjunction with local consistency, dynamically modulates the contribution of different components and pixels. This enables adaptive gradient backpropagation and allows the network to fully exploit its capacity. Compared to the state-of-the-art, our lightweight network, termed ArgMatch, achieves competitive performance in serval tasks, while significantly reducing the computational cost. Codes are available in https://github.com/ACuOoOoO/argmatch.