A cost effective estimation of depth from stereo image pairs using shallow siamese convolutional networks

Abstract

We propose a cost effective depth estimation method using stereo camera and convolutional neural networks for SLAM algorithm. Convolutional neural networks outperform the traditional computer vision approaches in estimating depth of stereo image pairs. However, the performance gain of neural networks approach causes substantial increase in computation cost, which consequently decreases the operation time of mobile robots like domestic robots. To alleviate the high computation problem, this paper proposes quantized shallow Siamese convolutional neural networks which compute the similarity between patches of rectified stereo image pairs to estimate depth. Quantization of weights and reduction of layers in the neural networks can degrade the performance. To mitigate the performance degradation, this paper initially maximizes networks performance with three different methods of batch-normalization, optimal negative matching similarity training, and retraining with a global loss function. Then, the final retrained network is nonuniformly quantized. This non-uniform quantization provides efficient computation with the minimum performance loss. The final quantized shallow Siamese networks achieve 3.29% error rate for KITTI 2012.