DMCNet: Toward Lightweight Volumetric Stereo Matching

Recent methods in stereo matching have steadily improved disparity prediction accuracy using volumetric deep convolutional neural network (CNN) architectures. However, this performance gain is achieved with a large memory footprint, making models from existing volumetric stereo networks unable to run on memory-constrained embedded vision devices. This challenge raises an important research question that inspires this work: Can we achieve a lightweight volumetric deep CNN architecture without sacrificing accuracy? To answer this question, we propose a discriminative multiscale context network (DMCNet), a memory-efficient backbone for volumetric stereo matching. Our key insight is to leverage a lightweight multibranch network to extract rich contextual information and a series of downsampling layers to achieve sufficient semantics with a large receptive field. In addition, we design a feature reuse and fusion (FRF) module that combines information from complementary scales to obtain a more robust representation for accurate disparity regression. We show that these network design recipes provide stronger representations with fewer parameters for learning high-quality disparity map prediction. Experiments on stereo benchmarks reveal that the proposed DMCNet achieves better parameter and memory efficiency as well as competitive accuracy compared to many existing state-of-the-art methods.