The unsupervised stereo matching algorithm GANet (Gated Aggregation Network) uses a deep neural network with a gating mechanism. To estimate the depth map from stereo picture pairs, it integrates cost volume construction, cost volume regularisation, and disparity regression.

A complete deep learning system for stereo matching is called PSMNet (Pyramid Stereo Matching Network). It makes use of a spatial pyramid pooling module to gather contextual data on several scales and a stacked hourglass design to gradually improve disparity estimation.

DeepPruner: DeepPruner is a deep learning-based unsupervised stereo-matching system that includes an occlusion-aware loss function. It tries to handle occlusions and boost the precision of disparity estimates.

Stereo matching is a computer vision task that involves finding correspondences between the pixels of two or more images taken from different viewpoints. End-to-end unsupervised learning algorithms for stereo matching aim to learn a mapping directly from the input images to the corresponding depth maps without relying on manually annotated ground truth data.

Here are a few notable end-to-end unsupervised learning algorithms for stereo matching:

These algorithms represent some of the advancements in end-to-end unsupervised learning for stereo matching. However, it's worth noting that research in this field is evolving rapidly, and new algorithms and techniques may emerge in the future.

Stereo matching algorithms are frequently used in computer vision to compare the discrepancies between two images collected from various angles in order to determine the depth information of a scene. Although there are several stereo matching algorithms, I'll concentrate on end-to-end unsupervised learning techniques because they don't need ground truth disparity maps for training. Please note that my understanding is based on data as of September 2021, and that there may have been developments or new algorithms after then.

1. End-to-end unsupervised stereo matching algorithm GANet (Generative Adversarial Network) uses a deep neural network architecture along with a generative adversarial network framework. A disparity estimation network and a discriminator network make up its two primary parts. The discriminator network makes a distinction between the predicted disparity maps and the ground truth disparities, while the disparity estimation network uses the left and right pictures as input to predict the appropriate disparity map. To enhance the accuracy of the calculated disparities, the two networks are trained in competition.

2. Another end-to-end unsupervised stereo matching approach that makes use of deep convolutional neural networks and gated warping is GwcNet (Gated Warping Convolution Network). In order to warp the right picture to the left view, it provides a soft warping module that trains to produce differentiable disparity maps. To estimate the discrepancies, the warped right image and the left image are combined and fed into a deep convolutional network. By reducing the photometric difference between the original left image and the reconstructed left image, the model is trained.

3. PSMNet: PSMNet (Pyramid Stereo Matching Network) is a multi-scale 3D convolutional neural network-based stereo matching technique that uses deep learning to estimate disparities. To extract multi-scale features, it uses a stacked hourglass network. To aggregate matching costs across various disparity levels, it uses a cost volume construction module. A 3D CNN module is used to refine the estimated disparities. In order to promote consistency between the left-to-right and right-to-left disparity estimations, PSMNet is trained using a self-supervised loss.

4. A deep neural network architecture is used by DeepPruner, an end-to-end unsupervised stereo matching technology, to evaluate disparities. With the addition of a cost volume pruning module, the algorithm becomes more effective by narrowing the search space for matching costs. When training, DeepPruner uses smoothness and left-right consistency constraints to regularize the disparity estimation procedure. Combining disparity smoothness loss and reconstruction loss is used to train the network.

The algorithms used in end-to-end unsupervised stereo matching are just a few examples. It's important to remember that computer vision and deep learning are fields that are continually changing, and that new methods may have been developed since my knowledge cutoff in September 2021. For the most recent knowledge on this subject, it is always a good idea to peruse the most recent study literature and publications.

This link has a useful search for you

Article Stereo Matching Algorithm based on Deep Learning: A Survey

End-to-end unsupervised learning stereo matching algorithms aim to estimate the depth or disparity maps from stereo image pairs without relying on ground truth depth or disparity annotations. These algorithms typically leverage deep learning techniques to learn the mapping directly from the input images. Here are a few notable approaches in this field: