Asymmetric cost aggregation network for efficient stereo matching

Cost aggregation is crucial to the accuracy of stereo matching. A reasonable cost algorithm should aggregate costs within homogeneous regions where pixels have same or similar disparities. Otherwise, estimated disparity map prone well-known edge-fattening issue and problem losing fine structures. However, current state-of-the-art convolutional neural networks (CNNs) mainly do with square-kernel layers that learn adjust their kernel elements make actual receptive fields aggregated adapt various shapes. This a mechanism easily leads above issues due translation equivalence content-insensitivity properties CNNs. To tackle these problems, novel densely connected asymmetric convolution block (Dense-ACB) based on convolutions proposed explicitly construct shapes, which effectively alleviates caused by mismatching shapes regions. The Dense-ACB brings new insight CNN-based matching networks. Based method, an efficient effective network built, not only achieves competitive overall compared models but also preserves