MFANet: Multi-feature Aggregation Network for Domain Generalized Stereo Matching

Stereo matching models based on Deep Neural Networks (DNNs) and trained on synthetic domains often struggle to generalize effectively to the imperceptible real domain. Presently, numerous stereo matching approaches focused on domain generalization solely involve a straightforward sequence of multi-scale semantic features. Unfortunately, this simplistic approach can lead to domain shift challenges owing to the omission of crucial detailed information. Simultaneously, this study asserts that texture features play a pivotal role in influencing domain generalization. This is due to the tendency of stereo matching algorithms to seek out the most dominant texture information for matching within texture-rich regions, often disregarding other critical data. To address this, this paper introduces adaptive semantic feature aggregation as well as multi-scale texture feature aggregation, effectively leveraging both the superficial texture data and profound semantic insights derived from the convolutional feature extraction network. This approach mitigates the risk of feature over-specialization while diminishing the influence of cross-domain disparities on the stereo matching network. Furthermore, the model incorporates Recurrent Neural Networks (RNNs) to instruct the cost aggregation network in information synthesis. The performance of our proposed MFANet in stereo matching attains the state-of-the-art level when trained on synthetic datasets and subsequently generalized to four distinct real datasets.