Use of CNNs for Estimating Depth from Stereo Images

In order to build disparity maps from stereo images, this research investigates the advantages of employing CNN (convolutional neural networks) to generate a disparity space image. A total of 3100 neurons are used to build an eight-layer fully connected network that is trained using 220,000 examples of both positive and negative image patch samples. Using a context-based technique, the disparity space image is aggregated. The WTA technique is used to create the disparity map. There are noticeable visual improvements as compared to the simple subtractive plane-sweep technique, particularly when there is little to no texture.