Dense-CNN: Dense convolutional neural network for stereo matching using multiscale feature connection

Highlights：

• We present a dense convolutional neural network with multiscale feature connection for stereo matching, named Dense-CNN. This proposed framework is able to capture rich context information to develop the performance of disparity estimation in ill-posed regions.

• We explore a novel loss-function strategy by employing the binary cross-entropy loss function with various expected values to supervise the training of the proposed Dense-CNN model. The proposed loss function is reasonably good in training network parameters.

• Experimental results of applications on the Middlebury and KITTI test datasets demonstrate that the proposed Dense-CNN method shows good performance on both accuracy and robustness of disparity estimation.