基于多尺度特征融合的单目图像深度估计

Monocular depth estimation with multi-scale feature fusion

为解决从单目图像中很难恢复出准确、有效深度信息的问题,提出一种多尺度特征融合的单目图像深度估计算法.算法采用端对端训练的卷积神经网络(CNN)结构,引入从图像编码器到解码器的跳层连接来实现在不同尺度上特征的提取和表达,设计了一种多尺度的损失函数来提升卷积神经网络的训练效果.通过在NYU Depth V2室内场景深度数据集和KITTI室外场景深度数据集上的训练、验证和测试,实验结果表明:提出的多尺度特征融合方法得到的深度图边缘清晰、层次分明,且在室内场景和室外场景中均能适用,具有较强的泛化性,可以适应多种实际场景的需求.

To solve the problem that it is difficult to recover accurate and effective depth information from monocular images,a monocular image depth estimation algorithm based on multi-scale feature fusion was proposed.End-to-end trained convolutional neural network(CNN)structure was applied to the algorithm,and the skip layer connection from image encoder to decoder was introduced to realize feature extraction and expression on different scales.A multi-scale loss function was designed to improve the training effect of the convolutional neural network.Through training,verification and testing on the NYU Depth V2 indoor scene depth dataset and KITTI outdoor scene depth dataset,experimental results show that the proposed multi-scale feature fusion method can obtain clear,sharp-edged edges in the depth map,and is applicable to both indoor and outdoor scenes with strong generalization,which can adapt to the demands of a variety of actual scenes.