We propose a novel learning-based method for multi-view stereo (MVS) depth estimation capable of recovering depth from images taken from known, but unconstrained, views. Existing MVS methods extract features from each image independently before projecting them onto a set of planes at candidate depths to compute matching costs. By projecting features after extraction, networks must learn rotation and scale invariant representations even though the relative poses of the cameras are known. In our approach, we compensate for viewpoint changes directly in the extraction layers, allowing the network to learn features that are projected by construction and reducing the need for rotation and scale invariance.Compensating for viewpoint changes naively, however, can be computationally expensive as the feature layers must either be applied multiple times (once per depth hypothesis), or replaced by 3D convolutions. We overcome this limitation in two ways. First, we only compute our matching cost volume at a coarse image scale before upsampling and refining the outputs. Second, we incrementally compute our projected features such that the bulk of the layers need only be executed a single time across all depth hypotheses. The combination of these two techniques allows our method to perform competitively with the state-of-the-art, while being significantly faster. We call our method MultiViewStereoNet and release our source code publicly for the benefit of the robotics community.

Although deep neural networks have achieved state-of-the-art performance for stereo depth estimation, they can suffer from a significant drop in accuracy when tested on images from novel domains. Recent work has shown that self-supervised online adaptation is a promising approach for closing this performance gap. In this work, we address three unsolved challenges for online adaptation. First, we propose a method for detecting novel environments, allowing us to trigger adaptation and notify downstream systems that depth predictions are unreliable. We find that the feature similarity scores from our deep stereo network can be leveraged for out-of-distribution (OOD) detection, providing the necessary starting criterion for adaptation. Next, we use online validation to terminate adaptation when it stops improving performance, allowing us to free up computational resources. Finally, we demonstrate that existing methods for continuous adaptation cause catastrophic forgetting of the training domain. By augmenting adaptation with experience replay, we retain high accuracy in the training domain while rapidly improving performance in novel environments. In sum, these three contributions form the basis of a more robust and efficient deep stereo system that can recognize and adapt to new environments without forgetting.

We propose an efficient method for monocular simultaneous localization and mapping (SLAM) that is capable of estimating metrically-scaled motion without additional sensors or hardware acceleration by integrating metric depth predictions from a neural network into a geometric SLAM factor graph. Unlike learned end-to-end SLAM systems, ours does not ignore the relative geometry directly observable in the images. Unlike existing learned depth estimation approaches, ours leverages the insight that when used to estimate scale, learned depth predictions need only be coarse in image space. This allows us to shrink our network to the point that performing inference on a standard CPU becomes computationally tractable.We make several improvements to our network architecture and training procedure to address the lack of depth observability when using coarse images, which allows us to estimate spatially coarse, but depth-accurate predictions in only 30 ms per frame without GPU acceleration. At runtime we incorporate the learned metric data as unary scale factors in a Sim(3) pose graph. Our method is able to generate accurate, scaled poses without additional sensors, hardware accelerators, or special maneuvers and does not ignore or corrupt the observable epipolar geometry. We show compelling results on the KITTI benchmark dataset in addition to real-world experiments with a handheld camera.

We propose a lightweight method for dense online monocular depth estimation capable of reconstructing 3D meshes on computationally constrained platforms. Our main contribution is to pose the reconstruction problem as a non-local variational optimization over a time-varying Delaunay graph of the scene geometry, which allows for an efficient, keyframeless approach to depth estimation. The graph can be tuned to favor reconstruction quality or speed and is continuously smoothed and augmented as the camera explores the scene. Unlike keyframe-based approaches, the optimized surface is always available at the current pose, which is necessary for low-latency obstacle avoidance. FLaME (Fast Lightweight Mesh Estimation) can generate mesh reconstructions at upwards of 230 Hz using less than one Intel i7 CPU core, which enables operation on size, weight, and power-constrained platforms. We present results from both benchmark datasets and experiments running FLaME in-the-loop onboard a small flying quadrotor.

We present a method for Simultaneous Localization and Mapping (SLAM) using a monocular camera that is capable of reconstructing dense 3D geometry online without the aid of a graphics processing unit (GPU). Our key contribution is a multi-resolution depth estimation and spatial smoothing process that exploits the correlation between low-texture image regions and simple planar structure to adaptively scale the complexity of the generated keyframe depthmaps to the texture of the input imagery. High-texture image regions are represented at higher resolutions to capture fine detail, while low-texture regions are represented at coarser resolutions for smooth surfaces. The computational savings enabled by this approach allow for significantly increased reconstruction density and quality when compared to the state-of-the-art. The increased depthmap density also improves tracking performance as more constraints can contribute to the pose estimation. A video of experimental results is available at http://groups.csail.mit.edu/rrg/multi_level_mapping.

We propose a method of real-time monocular camera-based localization in known environments. With the goal of controlling high-speed micro air vehicles (MAVs), we localize with respect to a mesh map of the environment that can support both pose estimation and trajectory planning. Using only limited hardware that can be carried on a MAV, we achieve accurate pose estimation at rates above 50 Hz, an order of magnitude faster than the current state-of-the-art mesh-based localization algorithms. In our simultaneous tracking and rendering (STAR) approach, we render virtual images of the environment and track camera images with respect to them using a robust semi-direct image alignment technique. Our main contribution is the decoupling of camera tracking from virtual image rendering, which drastically reduces the number of rendered images and enables accurate full camera-rate tracking without needing a high-end GPU. We demonstrate our approach in GPS-denied indoor environments.