Friedrich Fraundorfer

Dominik Hirner, Friedrich Fraundorfer

We propose an accurate and lightweight convolutional neural network for stereo estimation with depth completion. We name this method fully-convolutional deformable similarity network with depth completion (FCDSN-DC). This method extends FC-DCNN by improving the feature extractor, adding a network structure for training highly accurate similarity functions and a network structure for filling inconsistent disparity estimates. The whole method consists of three parts. The first part consists of fully-convolutional densely connected layers that computes expressive features of rectified image pairs. The second part of our network learns highly accurate similarity functions between this learned features. It consists of densely-connected convolution layers with a deformable convolution block at the end to further improve the accuracy of the results. After this step an initial disparity map is created and the left-right consistency check is performed in order to remove inconsistent points. The last part of the network then uses this input together with the corresponding left RGB image in order to train a network that fills in the missing measurements. Consistent depth estimations are gathered around invalid points and are parsed together with the RGB points into a shallow CNN network structure in order to recover the missing values. We evaluate our method on challenging real world indoor and outdoor scenes, in particular Middlebury, KITTI and ETH3D were it produces competitive results. We furthermore show that this method generalizes well and is well suited for many applications without the need of further training. The code of our full framework is available at: https://github.com/thedodo/FCDSN-DC

Ludwig Mohr, Ismail Geles, Friedrich Fraundorfer

Rigid registration of point clouds is a fundamental problem in computer vision with many applications from 3D scene reconstruction to geometry capture and robotics. If a suitable initial registration is available, conventional methods like ICP and its many variants can provide adequate solutions. In absence of a suitable initialization and in the presence of a high outlier rate or in the case of small overlap though the task of rigid registration still presents great challenges. The advent of deep learning in computer vision has brought new drive to research on this topic, since it provides the possibility to learn expressive feature-representations and provide one-shot estimates instead of depending on time-consuming iterations of conventional robust methods. Yet, the rotation and permutation invariant nature of point clouds poses its own challenges to deep learning, resulting in loss of performance and low generalization capability due to sensitivity to outliers and characteristics of 3D scans not present during network training. In this work, we present a novel fast and light-weight network architecture using the attention mechanism to augment point descriptors at inference time to optimally suit the registration task of the specific point clouds it is presented with. Employing a fully-connected graph both within and between point clouds lets the network reason about the importance and reliability of points for registration, making our approach robust to outliers, low overlap and unseen data. We test the performance of our registration algorithm on different registration and generalization tasks and provide information on runtime and resource consumption. The code and trained weights are available at https://github.com/mordecaimalignatius/GAFAR/.

Stefan Ainetter, Sinisa Stekovic, Friedrich Fraundorfer et al.

We present an automatic method for annotating images of indoor scenes with the CAD models of the objects by relying on RGB-D scans. Through a visual evaluation by 3D experts, we show that our method retrieves annotations that are at least as accurate as manual annotations, and can thus be used as ground truth without the burden of manually annotating 3D data. We do this using an analysis-by-synthesis approach, which compares renderings of the CAD models with the captured scene. We introduce a 'cloning procedure' that identifies objects that have the same geometry, to annotate these objects with the same CAD models. This allows us to obtain complete annotations for the ScanNet dataset and the recent ARKitScenes dataset.

Sinisa Stekovic, Mahdi Rad, Alireza Moradi et al.

We propose a novel method applicable in many scene understanding problems that adapts the Monte Carlo Tree Search (MCTS) algorithm, originally designed to learn to play games of high-state complexity. From a generated pool of proposals, our method jointly selects and optimizes proposals that minimize the objective term. In our first application for floor plan reconstruction from point clouds, our method selects and refines the room proposals, modelled as 2D polygons, by optimizing on an objective function combining the fitness as predicted by a deep network and regularizing terms on the room shapes. We also introduce a novel differentiable method for rendering the polygonal shapes of these proposals. Our evaluations on the recent and challenging Structured3D and Floor-SP datasets show significant improvements over the state-of-the-art, without imposing hard constraints nor assumptions on the floor plan configurations. In our second application, we extend our approach to reconstruct general 3D room layouts from a color image and obtain accurate room layouts. We also show that our differentiable renderer can easily be extended for rendering 3D planar polygons and polygon embeddings. Our method shows high performance on the Matterport3D-Layout dataset, without introducing hard constraints on room layout configurations.

Emanuele Santellani, Christian Sormann, Mattia Rossi et al.

In this work we introduce S-TREK, a novel local feature extractor that combines a deep keypoint detector, which is both translation and rotation equivariant by design, with a lightweight deep descriptor extractor. We train the S-TREK keypoint detector within a framework inspired by reinforcement learning, where we leverage a sequential procedure to maximize a reward directly related to keypoint repeatability. Our descriptor network is trained following a "detect, then describe" approach, where the descriptor loss is evaluated only at those locations where keypoints have been selected by the already trained detector. Extensive experiments on multiple benchmarks confirm the effectiveness of our proposed method, with S-TREK often outperforming other state-of-the-art methods in terms of repeatability and quality of the recovered poses, especially when dealing with in-plane rotations.

Emanuele Santellani, Christian Sormann, Mattia Rossi et al.

Establishing a sparse set of keypoint correspon dences between images is a fundamental task in many computer vision pipelines. Often, this translates into a computationally expensive nearest neighbor search, where every keypoint descriptor at one image must be compared with all the descriptors at the others. In order to lower the computational cost of the matching phase, we propose a deep feature extraction network capable of detecting a predefined number of complementary sets of keypoints at each image. Since only the descriptors within the same set need to be compared across the different images, the matching phase computational complexity decreases with the number of sets. We train our network to predict the keypoints and compute the corresponding descriptors jointly. In particular, in order to learn complementary sets of keypoints, we introduce a novel unsupervised loss which penalizes intersections among the different sets. Additionally, we propose a novel descriptor-based weighting scheme meant to penalize the detection of keypoints with non-discriminative descriptors. With extensive experiments we show that our feature extraction network, trained only on synthetically warped images and in a fully unsupervised manner, achieves competitive results on 3D reconstruction and re-localization tasks at a reduced matching complexity.

Christian Sormann, Emanuele Santellani, Mattia Rossi et al.

We propose a novel approach for deep learning-based Multi-View Stereo (MVS). For each pixel in the reference image, our method leverages a deep architecture to search for the corresponding point in the source image directly along the corresponding epipolar line. We denote our method DELS-MVS: Deep Epipolar Line Search Multi-View Stereo. Previous works in deep MVS select a range of interest within the depth space, discretize it, and sample the epipolar line according to the resulting depth values: this can result in an uneven scanning of the epipolar line, hence of the image space. Instead, our method works directly on the epipolar line: this guarantees an even scanning of the image space and avoids both the need to select a depth range of interest, which is often not known a priori and can vary dramatically from scene to scene, and the need for a suitable discretization of the depth space. In fact, our search is iterative, which avoids the building of a cost volume, costly both to store and to process. Finally, our method performs a robust geometry-aware fusion of the estimated depth maps, leveraging a confidence predicted alongside each depth. We test DELS-MVS on the ETH3D, Tanks and Temples and DTU benchmarks and achieve competitive results with respect to state-of-the-art approaches.

Stefan Ainetter, Sinisa Stekovic, Friedrich Fraundorfer et al.

We present an automated and efficient approach for retrieving high-quality CAD models of objects and their poses in a scene captured by a moving RGB-D camera. We first investigate various objective functions to measure similarity between a candidate CAD object model and the available data, and the best objective function appears to be a "render-and-compare" method comparing depth and mask rendering. We thus introduce a fast-search method that approximates an exhaustive search based on this objective function for simultaneously retrieving the object category, a CAD model, and the pose of an object given an approximate 3D bounding box. This method involves a search tree that organizes the CAD models and object properties including object category and pose for fast retrieval and an algorithm inspired by Monte Carlo Tree Search, that efficiently searches this tree. We show that this method retrieves CAD models that fit the real objects very well, with a speed-up factor of 10x to 120x compared to exhaustive search.

Emanuele Santellani, Martin Zach, Christian Sormann et al.

The extraction of keypoints in images is at the basis of many computer vision applications, from localization to 3D reconstruction. Keypoints come with a score permitting to rank them according to their quality. While learned keypoints often exhibit better properties than handcrafted ones, their scores are not easily interpretable, making it virtually impossible to compare the quality of individual keypoints across methods. We propose a framework that can refine, and at the same time characterize with an interpretable score, the keypoints extracted by any method. Our approach leverages a modified robust Gaussian Mixture Model fit designed to both reject non-robust keypoints and refine the remaining ones. Our score comprises two components: one relates to the probability of extracting the same keypoint in an image captured from another viewpoint, the other relates to the localization accuracy of the keypoint. These two interpretable components permit a comparison of individual keypoints extracted across different methods. Through extensive experiments we demonstrate that, when applied to popular keypoint detectors, our framework consistently improves the repeatability of keypoints as well as their performance in homography and two/multiple-view pose recovery tasks.

Tao Li, Zhenbao Yu, Banglei Guan et al.

This work presents two novel solvers for estimating the relative poses among views with known vertical directions. The vertical directions of camera views can be easily obtained using inertial measurement units (IMUs) which have been widely used in autonomous vehicles, mobile phones, and unmanned aerial vehicles (UAVs). Given the known vertical directions, our lgorithms only need to solve for two rotation angles and two translation vectors. In this paper, a linear closed-form solution has been described, requiring only four point correspondences in three views. We also propose a minimal solution with three point correspondences using the latest Gröbner basis solver. Since the proposed methods require fewer point correspondences, they can be efficiently applied within the RANSAC framework for outliers removal and pose estimation in visual odometry. The proposed method has been tested on both synthetic data and real-world scenes from KITTI. The experimental results show that the accuracy of the estimated poses is superior to other alternative methods.

Mattia D'Urso, Yuxi Hu, Christian Sormann et al.

Despite the growing need for data of more and more sophisticated 3D reconstruction pipelines, we can still observe a scarcity of suitable public datasets. Existing 3D datasets are either low resolution, limited to a small amount of scenes, based on images of varying quality because retrieved from the internet, or limited to specific capturing scenarios. Motivated by this lack of suitable 3D datasets, we captured TerraSky3D, a high-resolution large-scale 3D reconstruction dataset comprising 50,000 images divided into 150 ground, aerial, and mixed scenes. The dataset focuses on European landmarks and comes with curated calibration data, camera poses, and depth maps. TerraSky3D tries to answer the need for challenging dataset that can be used to train and evaluate 3D reconstruction-related pipelines.

Yuxi Hu, Jun Zhang, Kuangyi Chen et al.

Generalizable Gaussian Splatting aims to synthesize novel views for unseen scenes without per-scene optimization. In particular, recent advancements utilize feed-forward networks to predict per-pixel Gaussian parameters, enabling high-quality synthesis from sparse input views. However, existing approaches fall short in encoding discriminative, multi-view consistent features for Gaussian predictions, which struggle to construct accurate geometry with sparse views. To address this, we propose $\mathbf{C}^{3}$-GS, a framework that enhances feature learning by incorporating context-aware, cross-dimension, and cross-scale constraints. Our architecture integrates three lightweight modules into a unified rendering pipeline, improving feature fusion and enabling photorealistic synthesis without requiring additional supervision. Extensive experiments on benchmark datasets validate that $\mathbf{C}^{3}$-GS achieves state-of-the-art rendering quality and generalization ability. Code is available at: https://github.com/YuhsiHu/C3-GS.

Dominik Hirner, Friedrich Fraundorfer

Stereo estimation has made many advancements in recent years with the introduction of deep-learning. However the traditional supervised approach to deep-learning requires the creation of accurate and plentiful ground-truth data, which is expensive to create and not available in many situations. This is especially true for remote sensing applications, where there is an excess of available data without proper ground truth. To tackle this problem, we propose a self-supervised CNN with self-improving adaptive abilities. In the first iteration, the created disparity map is inaccurate and noisy. Leveraging the left-right consistency check, we get a sparse but more accurate disparity map which is used as an initial pseudo ground-truth. This pseudo ground-truth is then adapted and updated after every epoch in the training step of the network. We use the sum of inconsistent points in order to track the network convergence. The code for our method is publicly available at: https://github.com/thedodo/SAda-Net}{https://github.com/thedodo/SAda-Net

Stefano Zorzi, Shabab Bazrafkan, Stefan Habenschuss et al.

While most state-of-the-art instance segmentation methods produce binary segmentation masks, geographic and cartographic applications typically require precise vector polygons of extracted objects instead of rasterized output. This paper introduces PolyWorld, a neural network that directly extracts building vertices from an image and connects them correctly to create precise polygons. The model predicts the connection strength between each pair of vertices using a graph neural network and estimates the assignments by solving a differentiable optimal transport problem. Moreover, the vertex positions are optimized by minimizing a combined segmentation and polygonal angle difference loss. PolyWorld significantly outperforms the state of the art in building polygonization and achieves not only notable quantitative results, but also produces visually pleasing building polygons. Code and trained weights are publicly available at https://github.com/zorzi-s/PolyWorldPretrainedNetwork.

Dominik Hirner, Friedrich Fraundorfer

We propose a novel lightweight network for stereo estimation. Our network consists of a fully-convolutional densely connected neural network (FC-DCNN) that computes matching costs between rectified image pairs. Our FC-DCNN method learns expressive features and performs some simple but effective post-processing steps. The densely connected layer structure connects the output of each layer to the input of each subsequent layer. This network structure and the fact that we do not use any fully-connected layers or 3D convolutions leads to a very lightweight network. The output of this network is used in order to calculate matching costs and create a cost-volume. Instead of using time and memory-inefficient cost-aggregation methods such as semi-global matching or conditional random fields in order to improve the result, we rely on filtering techniques, namely median filter and guided filter. By computing a left-right consistency check we get rid of inconsistent values. Afterwards we use a watershed foreground-background segmentation on the disparity image with removed inconsistencies. This mask is then used to refine the final prediction. We show that our method works well for both challenging indoor and outdoor scenes by evaluating it on the Middlebury, KITTI and ETH3D benchmarks respectively. Our full framework is available at https://github.com/thedodo/FC-DCNN

Kuangyi Chen, Jun Zhang, Yuxi Hu et al.

Event cameras offer high-temporal-resolution sensing that remains reliable under high-speed motion and challenging lighting, making them promising for localization from LiDAR point clouds in GPS-denied and visually degraded environments. However, aligning sparse, asynchronous events with dense LiDAR maps is fundamentally ill-posed, as direct correspondence estimation suffers from modality gaps. We propose LEAR, a dual-task learning framework that jointly estimates edge structures and dense event-depth flow fields to bridge the sensing-modality divide. Instead of treating edges as a post-hoc aid, LEAR couples them with flow estimation through a cross-modal fusion mechanism that injects modality-invariant geometric cues into the motion representation, and an iterative refinement strategy that enforces mutual consistency between the two tasks over multiple update steps. This synergy produces edge-aware, depth-aligned flow fields that enable more robust and accurate pose recovery via Perspective-n-Point (PnP) solvers. On several popular and challenging datasets, LEAR achieves superior performance over the best prior method. The source code, trained models, and demo videos are made publicly available online.

Sinisa Stekovic, Arslan Artykov, Stefan Ainetter et al.

We propose PyTorchGeoNodes, a differentiable module for reconstructing 3D objects and their parameters from images using interpretable shape programs. Unlike traditional CAD model retrieval, shape programs allow reasoning about semantic parameters, editing, and a low memory footprint. Despite their potential, shape programs for 3D scene understanding have been largely overlooked. Our key contribution is enabling gradient-based optimization by parsing shape programs, or more precisely procedural models designed in Blender, into efficient PyTorch code. While there are many possible applications of our PyTochGeoNodes, we show that a combination of PyTorchGeoNodes with genetic algorithm is a method of choice to optimize both discrete and continuous shape program parameters for 3D reconstruction and understanding of 3D object parameters. Our modular framework can be further integrated with other reconstruction algorithms, and we demonstrate one such integration to enable procedural Gaussian splatting. Our experiments on the ScanNet dataset show that our method achieves accurate reconstructions while enabling, until now, unseen level of 3D scene understanding.

Yuxi Hu, Jun Zhang, Zhe Zhang et al.

Multi-view Stereo (MVS) aims to estimate depth and reconstruct 3D point clouds from a series of overlapping images. Recent learning-based MVS frameworks overlook the geometric information embedded in features and correlations, leading to weak cost matching. In this paper, we propose ICG-MVSNet, which explicitly integrates intra-view and cross-view relationships for depth estimation. Specifically, we develop an intra-view feature fusion module that leverages the feature coordinate correlations within a single image to enhance robust cost matching. Additionally, we introduce a lightweight cross-view aggregation module that efficiently utilizes the contextual information from volume correlations to guide regularization. Our method is evaluated on the DTU dataset and Tanks and Temples benchmark, consistently achieving competitive performance against state-of-the-art works, while requiring lower computational resources.

Kuangyi Chen, Jun Zhang, Friedrich Fraundorfer

Event cameras are bio-inspired sensors with some notable features, including high dynamic range and low latency, which makes them exceptionally suitable for perception in challenging scenarios such as high-speed motion and extreme lighting conditions. In this paper, we explore their potential for localization within pre-existing LiDAR maps, a critical task for applications that require precise navigation and mobile manipulation. Our framework follows a paradigm based on the refinement of an initial pose. Specifically, we first project LiDAR points into 2D space based on a rough initial pose to obtain depth maps, and then employ an optical flow estimation network to align events with LiDAR points in 2D space, followed by camera pose estimation using a PnP solver. To enhance geometric consistency between these two inherently different modalities, we develop a novel frame-based event representation that improves structural clarity. Additionally, given the varying degrees of bias observed in the ground truth poses, we design a module that predicts an auxiliary variable as a regularization term to mitigate the impact of this bias on network convergence. Experimental results on several public datasets demonstrate the effectiveness of our proposed method. To facilitate future research, both the code and the pre-trained models are made available online.

Mattia D'Urso, Emanuele Santellani, Christian Sormann et al.

We introduce SANDesc, a Streamlined Attention-Based Network for Descriptor extraction that aims to improve on existing architectures for keypoint description. Our descriptor network learns to compute descriptors that improve matching without modifying the underlying keypoint detector. We employ a revised U-Net-like architecture enhanced with Convolutional Block Attention Modules and residual paths, enabling effective local representation while maintaining computational efficiency. We refer to the building blocks of our model as Residual U-Net Blocks with Attention. The model is trained using a modified triplet loss in combination with a curriculum learning-inspired hard negative mining strategy, which improves training stability. Extensive experiments on HPatches, MegaDepth-1500, and the Image Matching Challenge 2021 show that training SANDesc on top of existing keypoint detectors leads to improved results on multiple matching tasks compared to the original keypoint descriptors. At the same time, SANDesc has a model complexity of just 2.4 million parameters. As a further contribution, we introduce a new urban dataset featuring 4K images and pre-calibrated intrinsics, designed to evaluate feature extractors. On this benchmark, SANDesc achieves substantial performance gains over the existing descriptors while operating with limited computational resources.

Yuchen Rao, Stefan Ainetter, Sinisa Stekovic et al.

High-level 3D scene understanding is essential in many applications. However, the challenges of generating accurate 3D annotations make development of deep learning models difficult. We turn to recent advancements in automatic retrieval of synthetic CAD models, and show that data generated by such methods can be used as high-quality ground truth for training supervised deep learning models. More exactly, we employ a pipeline akin to the one previously used to automatically annotate objects in ScanNet scenes with their 9D poses and CAD models. This time, we apply it to the recent ScanNet++ v1 dataset, which previously lacked such annotations. Our findings demonstrate that it is not only possible to train deep learning models on these automatically-obtained annotations but that the resulting models outperform those trained on manually annotated data. We validate this on two distinct tasks: point cloud completion and single-view CAD model retrieval and alignment. Our results underscore the potential of automatic 3D annotations to enhance model performance while significantly reducing annotation costs. To support future research in 3D scene understanding, we will release our annotations, which we call SCANnotate++, along with our trained models.

Christian Sormann, Mattia Rossi, Andreas Kuhn et al.

We present a novel deep-learning-based method for Multi-View Stereo. Our method estimates high resolution and highly precise depth maps iteratively, by traversing the continuous space of feasible depth values at each pixel in a binary decision fashion. The decision process leverages a deep-network architecture: this computes a pixelwise binary mask that establishes whether each pixel actual depth is in front or behind its current iteration individual depth hypothesis. Moreover, in order to handle occluded regions, at each iteration the results from different source images are fused using pixelwise weights estimated by a second network. Thanks to the adopted binary decision strategy, which permits an efficient exploration of the depth space, our method can handle high resolution images without trading resolution and precision. This sets it apart from most alternative learning-based Multi-View Stereo methods, where the explicit discretization of the depth space requires the processing of large cost volumes. We compare our method with state-of-the-art Multi-View Stereo methods on the DTU, Tanks and Temples and the challenging ETH3D benchmarks and show competitive results.

Stefan Ainetter, Christoph Böhm, Rohit Dhakate et al.

In this paper, we present a novel deep neural network architecture for joint class-agnostic object segmentation and grasp detection for robotic picking tasks using a parallel-plate gripper. We introduce depth-aware Coordinate Convolution (CoordConv), a method to increase accuracy for point proposal based object instance segmentation in complex scenes without adding any additional network parameters or computation complexity. Depth-aware CoordConv uses depth data to extract prior information about the location of an object to achieve highly accurate object instance segmentation. These resulting segmentation masks, combined with predicted grasp candidates, lead to a complete scene description for grasping using a parallel-plate gripper. We evaluate the accuracy of grasp detection and instance segmentation on challenging robotic picking datasets, namely Siléane and OCID_grasp, and show the benefit of joint grasp detection and segmentation on a real-world robotic picking task.

Stefan Ainetter, Friedrich Fraundorfer

In this work, we introduce a novel, end-to-end trainable CNN-based architecture to deliver high quality results for grasp detection suitable for a parallel-plate gripper, and semantic segmentation. Utilizing this, we propose a novel refinement module that takes advantage of previously calculated grasp detection and semantic segmentation and further increases grasp detection accuracy. Our proposed network delivers state-of-the-art accuracy on two popular grasp dataset, namely Cornell and Jacquard. As additional contribution, we provide a novel dataset extension for the OCID dataset, making it possible to evaluate grasp detection in highly challenging scenes. Using this dataset, we show that semantic segmentation can additionally be used to assign grasp candidates to object classes, which can be used to pick specific objects in the scene.

Sinisa Stekovic, Mahdi Rad, Friedrich Fraundorfer et al.

We propose a novel method for reconstructing floor plans from noisy 3D point clouds. Our main contribution is a principled approach that relies on the Monte Carlo Tree Search (MCTS) algorithm to maximize a suitable objective function efficiently despite the complexity of the problem. Like previous work, we first project the input point cloud to a top view to create a density map and extract room proposals from it. Our method selects and optimizes the polygonal shapes of these room proposals jointly to fit the density map and outputs an accurate vectorized floor map even for large complex scenes. To do this, we adapted MCTS, an algorithm originally designed to learn to play games, to select the room proposals by maximizing an objective function combining the fitness with the density map as predicted by a deep network and regularizing terms on the room shapes. We also introduce a refinement step to MCTS that adjusts the shape of the room proposals. For this step, we propose a novel differentiable method for rendering the polygonal shapes of these proposals. We evaluate our method on the recent and challenging Structured3D and Floor-SP datasets and show a significant improvement over the state-of-the-art, without imposing any hard constraints nor assumptions on the floor plan configurations.

Shreyas Hampali, Sinisa Stekovic, Sayan Deb Sarkar et al.

We explore how a general AI algorithm can be used for 3D scene understanding to reduce the need for training data. More exactly, we propose a modification of the Monte Carlo Tree Search (MCTS) algorithm to retrieve objects and room layouts from noisy RGB-D scans. While MCTS was developed as a game-playing algorithm, we show it can also be used for complex perception problems. Our adapted MCTS algorithm has few easy-to-tune hyperparameters and can optimise general losses. We use it to optimise the posterior probability of objects and room layout hypotheses given the RGB-D data. This results in an analysis-by-synthesis approach that explores the solution space by rendering the current solution and comparing it to the RGB-D observations. To perform this exploration even more efficiently, we propose simple changes to the standard MCTS' tree construction and exploration policy. We demonstrate our approach on the ScanNet dataset. Our method often retrieves configurations that are better than some manual annotations, especially on layouts.

Christian Sormann, Patrick Knöbelreiter, Andreas Kuhn et al.

In this work, we propose BP-MVSNet, a convolutional neural network (CNN)-based Multi-View-Stereo (MVS) method that uses a differentiable Conditional Random Field (CRF) layer for regularization. To this end, we propose to extend the BP layer and add what is necessary to successfully use it in the MVS setting. We therefore show how we can calculate a normalization based on the expected 3D error, which we can then use to normalize the label jumps in the CRF. This is required to make the BP layer invariant to different scales in the MVS setting. In order to also enable fractional label jumps, we propose a differentiable interpolation step, which we embed into the computation of the pairwise term. These extensions allow us to integrate the BP layer into a multi-scale MVS network, where we continuously improve a rough initial estimate until we get high quality depth maps as a result. We evaluate the proposed BP-MVSNet in an ablation study and conduct extensive experiments on the DTU, Tanks and Temples and ETH3D data sets. The experiments show that we can significantly outperform the baseline and achieve state-of-the-art results.

Stefano Zorzi, Ksenia Bittner, Friedrich Fraundorfer

We propose a machine learning based approach for automatic regularization and polygonization of building segmentation masks. Taking an image as input, we first predict building segmentation maps exploiting generic fully convolutional network (FCN). A generative adversarial network (GAN) is then involved to perform a regularization of building boundaries to make them more realistic, i.e., having more rectilinear outlines which construct right angles if required. This is achieved through the interplay between the discriminator which gives a probability of input image being true and generator that learns from discriminator's response to create more realistic images. Finally, we train the backbone convolutional neural network (CNN) which is adapted to predict sparse outcomes corresponding to building corners out of regularized building segmentation results. Experiments on three building segmentation datasets demonstrate that the proposed method is not only capable of obtaining accurate results, but also of producing visually pleasing building outlines parameterized as polygons.

Stefano Zorzi, Ksenia Bittner, Friedrich Fraundorfer

In the fast developing countries it is hard to trace new buildings construction or old structures destruction and, as a result, to keep the up-to-date cadastre maps. Moreover, due to the complexity of urban regions or inconsistency of data used for cadastre maps extraction, the errors in form of misalignment is a common problem. In this work, we propose an end-to-end deep learning approach which is able to solve inconsistencies between the input intensity image and the available building footprints by correcting label noises and, at the same time, misalignments if needed. The obtained results demonstrate the robustness of the proposed method to even severely misaligned examples that makes it potentially suitable for real applications, like OpenStreetMap correction.

Stefano Zorzi, Friedrich Fraundorfer

In this paper we present a method for building boundary refinement and regularization in satellite images using a fully convolutional neural network trained with a combination of adversarial and regularized losses. Compared to a pure Mask R-CNN model, the overall algorithm can achieve equivalent performance in terms of accuracy and completeness. However, unlike Mask R-CNN that produces irregular footprints, our framework generates regularized and visually pleasing building boundaries which are beneficial in many applications.

Banglei Guan, Ji Zhao, Daniel Barath et al.

We propose three novel solvers for estimating the relative pose of a multi-camera system from affine correspondences (ACs). A new constraint is derived interpreting the relationship of ACs and the generalized camera model. Using the constraint, we demonstrate efficient solvers for two types of motions assumed. Considering that the cameras undergo planar motion, we propose a minimal solution using a single AC and a solver with two ACs to overcome the degenerate case. Also, we propose a minimal solution using two ACs with known vertical direction, e.g., from an IMU. Since the proposed methods require significantly fewer correspondences than state-of-the-art algorithms, they can be efficiently used within RANSAC for outlier removal and initial motion estimation. The solvers are tested both on synthetic data and on real-world scenes from the KITTI odometry benchmark. It is shown that the accuracy of the estimated poses is superior to the state-of-the-art techniques.

Patrick Knöbelreiter, Christian Sormann, Alexander Shekhovtsov et al.

It has been proposed by many researchers that combining deep neural networks with graphical models can create more efficient and better regularized composite models. The main difficulties in implementing this in practice are associated with a discrepancy in suitable learning objectives as well as with the necessity of approximations for the inference. In this work we take one of the simplest inference methods, a truncated max-product Belief Propagation, and add what is necessary to make it a proper component of a deep learning model: We connect it to learning formulations with losses on marginals and compute the backprop operation. This BP-Layer can be used as the final or an intermediate block in convolutional neural networks (CNNs), allowing us to design a hierarchical model composing BP inference and CNNs at different scale levels. The model is applicable to a range of dense prediction problems, is well-trainable and provides parameter-efficient and robust solutions in stereo, optical flow and semantic segmentation.

Sinisa Stekovic, Shreyas Hampali, Mahdi Rad et al.

We present a novel method to reconstruct the 3D layout of a room (walls, floors, ceilings) from a single perspective view in challenging conditions, by contrast with previous single-view methods restricted to cuboid-shaped layouts. This input view can consist of a color image only, but considering a depth map results in a more accurate reconstruction. Our approach is formalized as solving a constrained discrete optimization problem to find the set of 3D polygons that constitute the layout. In order to deal with occlusions between components of the layout, which is a problem ignored by previous works, we introduce an analysis-by-synthesis method to iteratively refine the 3D layout estimate. As no dataset was available to evaluate our method quantitatively, we created one together with several appropriate metrics. Our dataset consists of 293 images from ScanNet, which we annotated with precise 3D layouts. It offers three times more samples than the popular NYUv2 303 benchmark, and a much larger variety of layouts.

Banglei Guan, Ji Zhao, Zhang Li et al.

In this paper we present four cases of minimal solutions for two-view relative pose estimation by exploiting the affine transformation between feature points and we demonstrate efficient solvers for these cases. It is shown, that under the planar motion assumption or with knowledge of a vertical direction, a single affine correspondence is sufficient to recover the relative camera pose. The four cases considered are two-view planar relative motion for calibrated cameras as a closed-form and a least-squares solution, a closed-form solution for unknown focal length and the case of a known vertical direction. These algorithms can be used efficiently for outlier detection within a RANSAC loop and for initial motion estimation. All the methods are evaluated on both synthetic data and real-world datasets from the KITTI benchmark. The experimental results demonstrate that our methods outperform comparable state-of-the-art methods in accuracy with the benefit of a reduced number of needed RANSAC iterations.

Andreas Kuhn, Christian Sormann, Mattia Rossi et al.

Deep Neural Networks (DNNs) have the potential to improve the quality of image-based 3D reconstructions. However, the use of DNNs in the context of 3D reconstruction from large and high-resolution image datasets is still an open challenge, due to memory and computational constraints. We propose a pipeline which takes advantage of DNNs to improve the quality of 3D reconstructions while being able to handle large and high-resolution datasets. In particular, we propose a confidence prediction network explicitly tailored for Multi-View Stereo (MVS) and we use it for both depth map outlier filtering and depth map refinement within our pipeline, in order to improve the quality of the final 3D reconstructions. We train our confidence prediction network on (semi-)dense ground truth depth maps from publicly available real world MVS datasets. With extensive experiments on popular benchmarks, we show that our overall pipeline can produce state-of-the-art 3D reconstructions, both qualitatively and quantitatively.

Sinisa Stekovic, Friedrich Fraundorfer, Vincent Lepetit

We propose a simple yet effective method to learn to segment new indoor scenes from video frames: State-of-the-art methods trained on one dataset, even as large as the SUNRGB-D dataset, can perform poorly when applied to images that are not part of the dataset, because of the dataset bias, a common phenomenon in computer vision. To make semantic segmentation more useful in practice, one can exploit geometric constraints. Our main contribution is to show that these constraints can be cast conveniently as semi-supervised terms, which enforce the fact that the same class should be predicted for the projections of the same 3D location in different images. This is interesting as we can exploit general existing techniques developed for semi-supervised learning to efficiently incorporate the constraints. We show that this approach can efficiently and accurately learn to segment target sequences of ScanNet and our own target sequences using only annotations from SUNRGB-D, and geometric relations between the video frames of target sequences.

Sinisa Stekovic, Friedrich Fraundorfer, Vincent Lepetit

We show that it is possible to learn semantic segmentation from very limited amounts of manual annotations, by enforcing geometric 3D constraints between multiple views. More exactly, image locations corresponding to the same physical 3D point should all have the same label. We show that introducing such constraints during learning is very effective, even when no manual label is available for a 3D point, and can be done simply by employing techniques from 'general' semi-supervised learning to the context of semantic segmentation. To demonstrate this idea, we use RGB-D image sequences of rigid scenes, for a 4-class segmentation problem derived from the ScanNet dataset. Starting from RGB-D sequences with a few annotated frames, we show that we can incorporate RGB-D sequences without any manual annotations to improve the performance, which makes our approach very convenient. Furthermore, we demonstrate our approach for semantic segmentation of objects on the LabelFusion dataset, where we show that one manually labeled image in a scene is sufficient for high performance on the whole scene.

Muhammad Shahzad, Michael Maurer, Friedrich Fraundorfer et al.

This paper addresses the highly challenging problem of automatically detecting man-made structures especially buildings in very high resolution (VHR) synthetic aperture radar (SAR) images. In this context, the paper has two major contributions: Firstly, it presents a novel and generic workflow that initially classifies the spaceborne TomoSAR point clouds $ - $ generated by processing VHR SAR image stacks using advanced interferometric techniques known as SAR tomography (TomoSAR) $ - $ into buildings and non-buildings with the aid of auxiliary information (i.e., either using openly available 2-D building footprints or adopting an optical image classification scheme) and later back project the extracted building points onto the SAR imaging coordinates to produce automatic large-scale benchmark labelled (buildings/non-buildings) SAR datasets. Secondly, these labelled datasets (i.e., building masks) have been utilized to construct and train the state-of-the-art deep Fully Convolution Neural Networks with an additional Conditional Random Field represented as a Recurrent Neural Network to detect building regions in a single VHR SAR image. Such a cascaded formation has been successfully employed in computer vision and remote sensing fields for optical image classification but, to our knowledge, has not been applied to SAR images. The results of the building detection are illustrated and validated over a TerraSAR-X VHR spotlight SAR image covering approximately 39 km$ ^2 $ $ - $ almost the whole city of Berlin $ - $ with mean pixel accuracies of around 93.84%

Georg Waltner, Michael Maurer, Thomas Holzmann et al.

Automated toll systems rely on proper classification of the passing vehicles. This is especially difficult when the images used for classification only cover parts of the vehicle. To obtain information about the whole vehicle. we reconstruct the vehicle as 3D object and exploit this additional information within a Convolutional Neural Network (CNN). However, when using deep networks for 3D object classification, large amounts of dense 3D models are required for good accuracy, which are often neither available nor feasible to process due to memory requirements. Therefore, in our method we reproject the 3D object onto the image plane using the reconstructed points, lines or both. We utilize this sparse depth prior within an auxiliary network branch that acts as a regularizer during training. We show that this auxiliary regularizer helps to improve accuracy compared to 2D classification on a real-world dataset. Furthermore due to the design of the network, at test time only the 2D camera images are required for classification which enables the usage in portable computer vision systems.

Tobias Koch, Lukas Liebel, Friedrich Fraundorfer et al.

While an increasing interest in deep models for single-image depth estimation methods can be observed, established schemes for their evaluation are still limited. We propose a set of novel quality criteria, allowing for a more detailed analysis by focusing on specific characteristics of depth maps. In particular, we address the preservation of edges and planar regions, depth consistency, and absolute distance accuracy. In order to employ these metrics to evaluate and compare state-of-the-art single-image depth estimation approaches, we provide a new high-quality RGB-D dataset. We used a DSLR camera together with a laser scanner to acquire high-resolution images and highly accurate depth maps. Experimental results show the validity of our proposed evaluation protocol.

Christian Mostegel, Friedrich Fraundorfer, Horst Bischof

In this work, we propose a novel approach to prioritize the depth map computation of multi-view stereo (MVS) to obtain compact 3D point clouds of high quality and completeness at low computational cost. Our prioritization approach operates before the MVS algorithm is executed and consists of two steps. In the first step, we aim to find a good set of matching partners for each view. In the second step, we rank the resulting view clusters (i.e. key views with matching partners) according to their impact on the fulfillment of desired quality parameters such as completeness, ground resolution and accuracy. Additional to geometric analysis, we use a novel machine learning technique for training a confidence predictor. The purpose of this confidence predictor is to estimate the chances of a successful depth reconstruction for each pixel in each image for one specific MVS algorithm based on the RGB images and the image constellation. The underlying machine learning technique does not require any ground truth or manually labeled data for training, but instead adapts ideas from depth map fusion for providing a supervision signal. The trained confidence predictor allows us to evaluate the quality of image constellations and their potential impact to the resulting 3D reconstruction and thus builds a solid foundation for our prioritization approach. In our experiments, we are thus able to reach more than 70% of the maximal reachable quality fulfillment using only 5% of the available images as key views. For evaluating our approach within and across different domains, we use two completely different scenarios, i.e. cultural heritage preservation and reconstruction of single family houses.

Xiao Xiang Zhu, Devis Tuia, Lichao Mou et al.

Standing at the paradigm shift towards data-intensive science, machine learning techniques are becoming increasingly important. In particular, as a major breakthrough in the field, deep learning has proven as an extremely powerful tool in many fields. Shall we embrace deep learning as the key to all? Or, should we resist a 'black-box' solution? There are controversial opinions in the remote sensing community. In this article, we analyze the challenges of using deep learning for remote sensing data analysis, review the recent advances, and provide resources to make deep learning in remote sensing ridiculously simple to start with. More importantly, we advocate remote sensing scientists to bring their expertise into deep learning, and use it as an implicit general model to tackle unprecedented large-scale influential challenges, such as climate change and urbanization.

Christian Häne, Lionel Heng, Gim Hee Lee et al.

Cameras are a crucial exteroceptive sensor for self-driving cars as they are low-cost and small, provide appearance information about the environment, and work in various weather conditions. They can be used for multiple purposes such as visual navigation and obstacle detection. We can use a surround multi-camera system to cover the full 360-degree field-of-view around the car. In this way, we avoid blind spots which can otherwise lead to accidents. To minimize the number of cameras needed for surround perception, we utilize fisheye cameras. Consequently, standard vision pipelines for 3D mapping, visual localization, obstacle detection, etc. need to be adapted to take full advantage of the availability of multiple cameras rather than treat each camera individually. In addition, processing of fisheye images has to be supported. In this paper, we describe the camera calibration and subsequent processing pipeline for multi-fisheye-camera systems developed as part of the V-Charge project. This project seeks to enable automated valet parking for self-driving cars. Our pipeline is able to precisely calibrate multi-camera systems, build sparse 3D maps for visual navigation, visually localize the car with respect to these maps, generate accurate dense maps, as well as detect obstacles based on real-time depth map extraction.

Christian Mostegel, Rudolf Prettenthaler, Friedrich Fraundorfer et al.

In this paper we present a scalable approach for robustly computing a 3D surface mesh from multi-scale multi-view stereo point clouds that can handle extreme jumps of point density (in our experiments three orders of magnitude). The backbone of our approach is a combination of octree data partitioning, local Delaunay tetrahedralization and graph cut optimization. Graph cut optimization is used twice, once to extract surface hypotheses from local Delaunay tetrahedralizations and once to merge overlapping surface hypotheses even when the local tetrahedralizations do not share the same topology.This formulation allows us to obtain a constant memory consumption per sub-problem while at the same time retaining the density independent interpolation properties of the Delaunay-based optimization. On multiple public datasets, we demonstrate that our approach is highly competitive with the state-of-the-art in terms of accuracy, completeness and outlier resilience. Further, we demonstrate the multi-scale potential of our approach by processing a newly recorded dataset with 2 billion points and a point density variation of more than four orders of magnitude - requiring less than 9GB of RAM per process.

Christian Mostegel, Markus Rumpler, Friedrich Fraundorfer et al.

In this paper we present an autonomous system for acquiring close-range high-resolution images that maximize the quality of a later-on 3D reconstruction with respect to coverage, ground resolution and 3D uncertainty. In contrast to previous work, our system uses the already acquired images to predict the confidence in the output of a dense multi-view stereo approach without executing it. This confidence encodes the likelihood of a successful reconstruction with respect to the observed scene and potential camera constellations. Our prediction module runs in real-time and can be trained without any externally recorded ground truth. We use the confidence prediction for on-site quality assurance and for planning further views that are tailored for a specific multi-view stereo approach with respect to the given scene. We demonstrate the capabilities of our approach with an autonomous Unmanned Aerial Vehicle (UAV) in a challenging outdoor scenario.

Christian Mostegel, Markus Rumpler, Friedrich Fraundorfer et al.

Learned confidence measures gain increasing importance for outlier removal and quality improvement in stereo vision. However, acquiring the necessary training data is typically a tedious and time consuming task that involves manual interaction, active sensing devices and/or synthetic scenes. To overcome this problem, we propose a new, flexible, and scalable way for generating training data that only requires a set of stereo images as input. The key idea of our approach is to use different view points for reasoning about contradictions and consistencies between multiple depth maps generated with the same stereo algorithm. This enables us to generate a huge amount of training data in a fully automated manner. Among other experiments, we demonstrate the potential of our approach by boosting the performance of three learned confidence measures on the KITTI2012 dataset by simply training them on a vast amount of automatically generated training data rather than a limited amount of laser ground truth data.