ROSSINI: Reconstructing 3D structure from single images: a perceptual reconstruction approach

消费者在电影、电视和虚拟现实（VR）中享受3D内容的沉浸式体验，但制作成本很高。拍摄3D电影需要两台摄像机来模拟观众的两只眼睛。一种常见但昂贵的替代方法是拍摄单个视图，然后在后期制作中使用视频艺术家来创建左右眼睛的视图。如果计算机可以从2D内容自动生成3D模型（和双目图像）：“将图像提升到3D”会怎么样？这是这个项目的首要目标。升降进入3D有多种用途，例如机器人的路线规划，自动驾驶汽车的避障，以及VR和电影中的应用。从2D图像估计3D结构是困难的，因为原则上，图像可能是由无数个3D场景创建的。识别这些可能的世界中哪一个是正确的是非常困难的，然而人类总是将2D图像解释为3D场景。每当我们看照片、看电视或凝视远方时，我们都会这样做，因为双眼深度线索很弱。虽然我们在判断距离时犯了一些错误，但我们快速理解任何场景布局的能力使我们能够在任何环境中导航并与之互动。计算机科学家已经建立了通过结合场景约束来提升到3D的机器视觉系统。一种流行的技术是用2D图像和相关的3D距离数据集训练深度神经网络。然而，为了取得成功，这种方法需要一个非常大的数据集，而获取数据集的成本可能很高。此外，只有当数据集完整时，性能才会好：如果系统遇到不符合训练数据集的场景或几何类型，它就会失败。大多数方法都是针对特定情况（例如室内或街道场景）进行训练的，这些系统对于农村场景的效果通常较差，并且不如人类灵活和健壮。最后，这样的系统提供了一个单一的重建输出，没有任何测量的不确定性。用户必须假设三维重建是正确的，在许多情况下，这将是一个昂贵的假设。计算机系统是根据其相对于现实世界的准确性来设计和评估的。然而，提升到3D的最终目标并不是完美的精度，而是提供一个3D表示，为人类观察者提供一个有用的和引人注目的视觉体验，或者引导机器人避开障碍物。重要的是，人类是与3D环境互动的专家，尽管我们的感知可能会大大偏离真实的度量深度。这表明，在任何环境中，类似人类的表征都是可以实现的，而且是足够的。ROSSINI将开发一种新的3D重建机器视觉系统，该系统比以前的方法更加灵活和健壮。专注于静态图像，我们将确定对人类重要的关键结构特征。我们将神经网络与计算机视觉方法相结合，形成类人的场景描述和3D场景模型。我们的目标是(i)产生对人类来说正确的3D表示，即使它们不是严格的几何正确（ii）对所有类型的场景都这样做，（iii）表达每次重建中固有的不确定性。为此，我们将收集人类解读图像的数据，并将这些信息整合到我们的网络中。我们的新训练方法将从人类和现有的地面真实数据集学习；训练算法选择最有用的人类任务（即在特定图像中判断深度）来最大化学习。重要的是，包含人类感知数据应该减少所需训练数据的总体数量，同时降低过度依赖特定数据集的风险。此外，经过充分训练后，我们的系统将生成3D重建以及深度估计的可靠性信息。

Consumers enjoy the immersive experience of 3D content in cinema, TV and virtual reality (VR), but it is expensive to produce. Filming a 3D movie requires two cameras to simulate the two eyes of the viewer. A common but expensive alternative is to film a single view, then use video artists to create the left and right eyes' views in post-production. What if a computer could automatically produce a 3D model (and binocular images) from 2D content: 'lifting images into 3D'? This is the overarching aim of this project. Lifting into 3D has multiple uses, such as route planning for robots, obstacle avoidance for autonomous vehicles, alongside applications in VR and cinema.Estimating 3D structure from a 2D image is difficult because in principle, the image could have been created from an infinite number of 3D scenes. Identifying which of these possible worlds is correct is very hard, yet humans interpret 2D images as 3D scenes all the time. We do this every time we look at a photograph, watch TV or gaze into the distance, where binocular depth cues are weak. Although we make some errors in judging distances, our ability to quickly understand the layout of any scene enables us to navigate through and interact with any environment.Computer scientists have built machine vision systems for lifting to 3D by incorporating scene constraints. A popular technique is to train a deep neural network with a collection of 2D images and associated 3D range data. However, to be successful, this approach requires a very large dataset, which can be expensive to acquire. Furthermore, performance is only as good as the dataset is complete: if the system encounters a type of scene or geometry that does not conform to the training dataset, it will fail. Most methods have been trained for specific situations - e.g. indoor, or street scenes - and these systems are typically less effective for rural scenes and less flexible and robust than humans. Finally, such systems provide a single reconstructed output, without any measure of uncertainty. The user must assume that the 3D reconstruction is correct, which will be a costly assumption in many cases.Computer systems are designed and evaluated based upon their accuracy with respect to the real world. However, the ultimate goal of lifting into 3D is not perfect accuracy - rather it is to deliver a 3D representation that provides a useful and compelling visual experience for a human observer, or to guide a robot whilst avoiding obstacles. Importantly, humans are expert at interacting with 3D environments, even though our perception can deviate substantially from true metric depth. This suggests that human-like representations are both achievable and sufficient, in any and all environments.ROSSINI will develop a new machine vision system for 3D reconstruction that is more flexible and robust than previous methods. Focussing on static images, we will identify key structural features that are important to humans. We will combine neural networks with computer vision methods to form human-like descriptions of scenes and 3D scene models. Our aims are to (i) produce 3D representations that look correct to humans even if they are not strictly geometrically correct (ii) do so for all types of scene and (iii) express the uncertainty inherent in each reconstruction. To this end we will collect data on human interpretation of images and incorporate this information into our network. Our novel training method will learn from humans and existing ground truth datasets; the training algorithm selecting the most useful human tasks (i.e. judge depth within a particular image) to maximise learning. Importantly, the inclusion of human perceptual data should reduce the overall quantity of training data required, while mitigating the risk of over-reliance on a specific dataset. Moreover, when fully trained, our system will produce 3D reconstructions alongside information about the reliability of the depth estimates.