When using deep stereo networks to predict disparity maps in real-world scenes, the network's accuracy tends to decline. This is due to the differences between dataset images and actual scene images. To enhance the network's performance in real-world scenarios, fine-tuning of the network parameters is necessary. In practice, underwater images can be easily obtained, but the corresponding disparity labels for training the network are difficult to acquire. This paper employs an unsupervised learning approach to fine-tune the network  using underwater images, allowing the stereo-matching network to achieve better performance in underwater environments. The network is applied to the measurement of abalone. 

stereo matching, deep learning, unsupervised learning, underwater measurement