М.С. Романов

Nikolay Patakin, Mikhail Romanov, Anna Vorontsova et al.

Nowadays, robotics, AR, and 3D modeling applications attract considerable attention to single-view depth estimation (SVDE) as it allows estimating scene geometry from a single RGB image. Recent works have demonstrated that the accuracy of an SVDE method hugely depends on the diversity and volume of the training data. However, RGB-D datasets obtained via depth capturing or 3D reconstruction are typically small, synthetic datasets are not photorealistic enough, and all these datasets lack diversity. The large-scale and diverse data can be sourced from stereo images or stereo videos from the web. Typically being uncalibrated, stereo data provides disparities up to unknown shift (geometrically incomplete data), so stereo-trained SVDE methods cannot recover 3D geometry. It was recently shown that the distorted point clouds obtained with a stereo-trained SVDE method can be corrected with additional point cloud modules (PCM) separately trained on the geometrically complete data. On the contrary, we propose GP$^{2}$, General-Purpose and Geometry-Preserving training scheme, and show that conventional SVDE models can learn correct shifts themselves without any post-processing, benefiting from using stereo data even in the geometry-preserving setting. Through experiments on different dataset mixtures, we prove that GP$^{2}$-trained models outperform methods relying on PCM in both accuracy and speed, and report the state-of-the-art results in the general-purpose geometry-preserving SVDE. Moreover, we show that SVDE models can learn to predict geometrically correct depth even when geometrically complete data comprises the minor part of the training set.

Mikhail Romanov, Nikolay Patatkin, Anna Vorontsova et al.

Single-view depth estimation (SVDE) plays a crucial role in scene understanding for AR applications, 3D modeling, and robotics, providing the geometry of a scene based on a single image. Recent works have shown that a successful solution strongly relies on the diversity and volume of training data. This data can be sourced from stereo movies and photos. However, they do not provide geometrically complete depth maps (as disparities contain unknown shift value). Therefore, existing models trained on this data are not able to recover correct 3D representations. Our work shows that a model trained on this data along with conventional datasets can gain accuracy while predicting correct scene geometry. Surprisingly, only a small portion of geometrically correct depth maps are required to train a model that performs equally to a model trained on the full geometrically correct dataset. After that, we train computationally efficient models on a mixture of datasets using the proposed method. Through quantitative comparison on completely unseen datasets and qualitative comparison of 3D point clouds, we show that our model defines the new state of the art in general-purpose SVDE.

Danil Galeev, Konstantin Sofiiuk, Danila Rukhovich et al.

In recent years generative models of visual data have made a great progress, and now they are able to produce images of high quality and diversity. In this work we study representations learnt by a GAN generator. First, we show that these representations can be easily projected onto semantic segmentation map using a lightweight decoder. We find that such semantic projection can be learnt from just a few annotated images. Based on this finding, we propose LayerMatch scheme for approximating the representation of a GAN generator that can be used for unsupervised domain-specific pretraining. We consider the semi-supervised learning scenario when a small amount of labeled data is available along with a large unlabeled dataset from the same domain. We find that the use of LayerMatch-pretrained backbone leads to superior accuracy compared to standard supervised pretraining on ImageNet. Moreover, this simple approach also outperforms recent semi-supervised semantic segmentation methods that use both labeled and unlabeled data during training. Source code for reproducing our experiments will be available at the time of publication.