On Regression Losses for Deep Depth Estimation

Abstract

Depth estimation from a single monocular image has reached great performances thanks to recent works based on deep networks. However, as various choices of losses, architectures and experimental conditions are proposed in the literature, it is difficult to establish their respective influence on the performances. In this paper we propose an in-depth study of various losses and experimental conditions for depth regression, on NYUv2 dataset. From this study we propose a new network for depth estimation combining an encoder-decoder architecture with an adversarial loss. This network reaches top ones state of the art on NUYv2 dataset while being simpler to train in a single phase.

Deep Depth from Defocus: How Can Defocus Blur Improve 3D Estimation Using Dense Neural Networks?

“…The NYU-Depth V2 (NYUv2) dataset [35] has approximately 230k pairs of images from 249 scenes for training and 215 scenes for testing. In [16], D3-Net reaches its best performances when trained with the complete dataset. However, NYUv2 also contains a smaller split with 1449 pairs of aligned RGB and depth images, of which 795 pairs are used for training and 654 pairs for testing.…”

“…To perform such tests, we adopt the D3-Net architecture from [16], illustrated in figure 3. We use the PyTorch framework on a NVIDIA TITAN X GPU with 12GB of memory.…”

“…To adapt the network, we replace the objective function of D3-Net by the one proposed by the authors of the dataset [17]. Then, for training, we take the weights of D3-Net trained on all-in-focus NYUv2 [16], and finetune the model on DiW using the modified network. We show the results of this model on the test set of DiW in figure 12.…”

“…In this paper, we use a dense neural network, D3-Net [16], in order to study the influence of defocus blur on depth estimation. Depth estimation performance is first tested on a synthetically defocused dataset created from NYUv2, with optically realistic blur variation, which allows to compare several optical settings and study their influence.…”

Estimação monocular de profundidade por aprendizagem profunda para veículos autônomos: influência da esparsidade dos mapas de profundidade no treinamento supervisionado

“…A seguir estão apresentadas as funções de custo que foram utilizadas para aprendizado supervisionado, no caso a supervisão é feita por mapas de profundidade. Mais especificamente, empregou-se as funções de custo mse, eigen e berhu para ajustar os parâmetros internos da rede neural profunda apresentada (CARVALHO et al, 2018). A motivação por trás disso foi simplesmente determinar qual seria a função mais adequada.…”