Template matching has been shown to accurately estimate the pose of a new object given a limited number of samples. However, pose estimation of occluded objects is still challenging. Furthermore, many robot application domains encounter texture-less objects for which depth images are more suitable than color images. In this paper, we propose a novel framework, Multi-Task Template Matching (MTTM), that finds the nearest template of a target object from a depth image while predicting segmentation masks and a pose transformation between the template and a detected object in the scene using the same feature map of the object region. The proposed feature comparison network computes segmentation masks and pose predictions by comparing feature maps of templates and cropped features of a scene. The segmentation result from this network improves the robustness of the pose estimation by excluding points that do not belong to the object. Experimental results show that MTTM outperforms baseline methods for segmentation and pose estimation of occluded objects despite using only depth images.

Dense 3D reconstructions generate globally consistent data of the environment suitable for many robot applications. Current RGB-D based reconstructions, however, only maintain the color resolution equal to the depth resolution of the used sensor. This firmly limits the precision and realism of the generated reconstructions. In this paper we present a real-time approach for creating and maintaining a surface reconstruction in as high as possible geometrical fidelity with full sensor resolution for its colorization (or surface texture). A multi-scale memory management process and a Level of Detail scheme enable equally detailed reconstructions to be generated at small scales, such as objects, as well as large scales, such as rooms or buildings. We showcase the benefit of this novel pipeline with a PrimeSense RGB-D camera as well as combining the depth channel of this camera with a high resolution global shutter camera. Further experiments show that our memory management approach allows us to scale up to larger domains that are not achievable with current state-of-the-art methods.

Applications that provide location related services need to understand the environment in which humans live such that verbal references and human interaction are possible. We formulate this semantic labelling task as the problem of learning the semantic labels from the perceived 3D structure. In this contribution we propose a batch approach and a novel multi-view frame fusion technique to exploit multiple views for improving the semantic labelling results. The batch approach works offline and is the direct application of an existing single-view method to scene reconstructions with multiple views. The multi-view frame fusion works in an incremental fashion accumulating the single-view results, hence allowing the online multi-view semantic segmentation of single frames and the offline reconstruction of semantic maps. Our experiments show the superiority of the approaches based on our fusion scheme, which leads to a more accurate semantic labelling.

In the last few years novel color and depth (RGB-D) sensors have greatly pushed robot perception. To enable a precise pixel-wise fusion of color and depth information good calibration is needed. The calibration determines the intrinsic parameters, the extrinsic parameters, and corrects for depth errors. While classic calibration approaches involve a dedicated calibration target and a trained expert, the autonomous calibration of such camera systems for robots operating in unknown environments is still an open problem. It demands for robust methods that do not need an expert to set up or tune the algorithm. Hence, we present a robust calibration algorithm that utilizes structure from motion (SfM) reconstructions as a calibration target and incorporates plane priors in the optimization to improve the convergence behavior and improve the calibration robustness. We evaluate our method against the state of the art performing over 300 experiments on ten different datasets, and show a significant improvement of the calibration accuracy.

The research in dense online 3D mapping is mostly focused on the geometrical accuracy and spatial extent of the reconstructions. Their color appearance is often neglected, leading to inconsistent colors and noticeable artifacts. We rectify this by extending a state-of-the-art SLAM system to accumulate colors in HDR space. We replace the simplistic pixel intensity averaging scheme with HDR color fusion rules tailored to the incremental nature of SLAM and a noise model suitable for off-the-shelf RGB-D cameras. Our main contribution is a map-aware exposure time controller. It makes decisions based on the global state of the map and predicted camera motion, attempting to maximize the information gain of each observation. We report a set of experiments demonstrating the improved texture quality and advantages of using the custom controller that is tightly integrated in the mapping loop.

In this paper we propose a multi-modal object recognition system that uses a two-step hypothesis verification approach to improve runtime efficiency. The system uses local and global appearance and shape features, generating many possibly competing hypotheses, which are then verified such that the scene can be optimally explained in terms of recognized object models. The introduced modification in this time consuming step reduces runtime considerably, while maintaining recognition performance. We evaluate recognition performance for various feature extraction modalities on the publicly available Willow Garage RGB-D dataset and show runtime improvements of a factor 2 to 10.

Cheap RGB-D sensors are ubiquitous in robotics. They typically contain a consumer-grade color camera that suffers from significant optical nonlinearities, often referred to as vignetting effects. For example, in Asus Xtion Live Pro cameras the pixels in the corners are two times darker than those in the center of the image. This deteriorates the visual appearance of 3D maps built with such cameras. We propose a simple calibration method that only requires a sheet of white paper as a calibration object and allows to reliably recover the vignetting response of a camera. We demonstrate calibration results for multiple popular RGB-D sensors and show that removal of vignetting effects using a nonparametric response model results in improved color coherence of the reconstructed maps. Furthermore, we show how to effectively compensate color variations caused by automatic white balance and exposure time control of the camera.

In this paper, we present an efficient semantic segmentation framework for indoor scenes operating on 3D point clouds. We use the results of a Random Forest Classifier to initialize the unary potentials of a densely interconnected Conditional Random Field, for which we learn the parameters for the pairwise potentials from training data. These potentials capture and model common spatial relations between class labels, which can often be observed in indoor scenes. We evaluate our approach on the popular NYU Depth datasets, for which it achieves superior results compared to the current state of the art. Exploiting parallelization and applying an efficient CRF inference method based on mean field approximation, our framework is able to process full resolution Kinect point clouds in half a second on a regular laptop, more than twice as fast as comparable methods.

This work presents a flexible system to reconstruct 3D models of objects captured with an RGB-D sensor. A major advantage of the method is that unlike other modelling tools, our reconstruction pipeline allows the user to acquire a full 3D model of the object. This is achieved by acquiring several partial 3D models in different sessions-each individual session presenting the object of interest in different configurations that reveal occluded parts of the object - that are automatically merged together to reconstruct a full 3D model. In addition, the 3D models acquired by our system can be directly used by state-of-the-art object instance recognition and object tracking modules, providing object-perception capabilities to complex applications requiring these functionalities (e.g. human-object interaction analysis, robot grasping, etc.). The system does not impose constraints in the appearance of objects (textured, untextured) nor in the modelling setup (moving camera with static object or turn-table setups with static camera). The proposed reconstruction system has been used to model a large number of objects resulting in metrically accurate and visually appealing 3D models.

Object segmentation of unknown objects with arbitrary shape in cluttered scenes is an ambitious goal in computer vision and became a great impulse with the introduction of cheap and powerful RGB-D sensors. We introduce a framework for segmenting RGB-D images where data is processed in a hierarchical fashion. After pre-clustering on pixel level parametric surface patches are estimated. Different relations between patch-pairs are calculated, which we derive from perceptual grouping principles, and support vector machine classification is employed to learn Perceptual Grouping. Finally, we show that object hypotheses generation with Graph-Cut finds a globally optimal solution and prevents wrong grouping. Our framework is able to segment objects, even if they are stacked or jumbled in cluttered scenes. We also tackle the problem of segmenting objects when they are partially occluded. The work is evaluated on publicly available object segmentation databases and also compared with state-of-the-art work of object segmentation.

This paper proposes an effective algorithm for recognizing objects and accurately estimating their 6DOF pose in scenes acquired by a RGB-D sensor. The proposed method is based on a combination of different recognition pipelines, each exploiting the data in a diverse manner and generating object hypotheses that are ultimately fused together in an Hypothesis Verification stage that globally enforces geometrical consistency between model hypotheses and the scene. Such a scheme boosts the overall recognition performance as it enhances the strength of the different recognition pipelines while diminishing the impact of their specific weaknesses. The proposed method outperforms the state-of-the-art on two challenging benchmark datasets for object recognition comprising 35 object models and, respectively, 176 and 353 scenes.

With the availability of cheap and powerful RGB-D sensors interest in 3D point cloud based methods has drastically increased. One common prerequisite of these methods is to abstract away from raw point cloud data, e.g. to planar patches, to reduce the amount of data and to handle noise and clutter. We present a novel method to abstract RGB-D sensor data to parametric surface models described by B-spline surfaces and associated boundaries. Data is first pre-segmented into smooth patches before B-spline surfaces are fitted. The best surface representations of these patches are selected in a merging procedure. Furthermore, we show how curve fitting estimates smooth boundaries and improves the given sensor information compared to hand-labelled ground truth annotation when using colour in addition to depth information. All parts of the framework are open-source1and are evaluated on the object segmentation database (OSD) also available online, showing accuracy and usability of the proposed methods.

Detecting elements such as planes in 3D is essential to describe objects for applications such as robotics and augmented reality. While plane estimation is well studied, table-top scenes exhibit a large number of planes and methods often lock onto a dominant plane or do not estimate 3D object structure but only homographies of individual planes. In this paper we introduce MDL to the problem of incrementally detecting multiple planar patches in a scene using tracked interest points in image sequences. Planar patches are reconstructed and stored in a keyframe-based graph structure. In case different motions occur, separate object hypotheses are modelled from currently visible patches and patches seen in previous frames. We evaluate our approach on a standard data set published by the Visual Geometry Group at the University of Oxford [24] and on our own data set containing table-top scenes. Results indicate that our approach significantly improves over the state-of-the-art algorithms.

We present a framework for segmenting unknown objects in RGB-D images suitable for robotics tasks such as object search, grasping and manipulation. While handling single objects on a table is solved, handling complex scenes poses considerable problems due to clutter and occlusion. After pre-segmentation of the input image based on surface normals, surface patches are estimated using a mixture of planes and NURBS (non-uniform rational B-splines) and model selection is employed to find the best representation for the given data. We then construct a graph from surface patches and relations between pairs of patches and perform graph cut to arrive at object hypotheses segmented from the scene. The energy terms for patch relations are learned from user annotated training data, where support vector machines (SVM) are trained to classify a relation as being indicative of two patches belonging to the same object. We show evaluation of the relations and results on a database of different test sets, demonstrating that the approach can segment objects of various shapes in cluttered table top scenes.

Man-made environments are abundant with planar surfaces which have attractive properties for robotics manipulation tasks and are a prerequisite for a variety of vision tasks. This work presents automatic on-line 3D object model acquisition assuming a robot to manipulate the object. Objects are represented with piecewise planar surfaces in a spatio-temporal graph. Planes once detected as homographies are tracked and serve as priors in subsequent images. After reconstruction of the planes the 3D motion is analyzed and initial object hypotheses are created. In case planes start moving independently a split event is triggered, the spatio-temporal object graph is traced back and visible planes as well as occluded planes are assigned to the most probable split object. The novelty of this framework is to formalize Multi-body Structure-and-Motion (MSaM), that is, to segment interest point tracks into different rigid objects and compute the multiple-view geometry of each object, with Minimal Description Length (MDL) based on model selection of planes in an incremental manner. Thus, object models are built from planes, which directly can be used for robotic manipulation.

Allowing a robot to acquire 3D object models autonomously not only requires robust feature detection and learning methods but also mechanisms for guiding learning and assessing learning progress. In this paper we present probabilistic measures for observed detection success, predicted detection success and the completeness of learned models, where learning is incremental and online. This allows the robot to decide when to add a new keyframe to its view-based object model, where to look next in order to complete the model, predicting the probability of successful object detection given the model trained so far as well as knowing when to stop learning.

Man-made environments are abundant with planar surfaces which have attractive properties and are a prerequisite for a variety of vision tasks. This paper presents an incremental model selection method to detect piecewise planar surfaces, where planes once detected are tracked and serve as priors in subsequent images. The novelty of this approach is to formalize model selection for plane detection with Minimal Description Length (MDL) in an incremental manner. In each iteration tracked planes and new planes computed from randomly sampled interest points are evaluated, the hypotheses which best explain the scene are retained, and their supporting points are marked so that in the next iteration random sampling is guided to unexplained points. Hence, the remaining finer scene details can be represented. We show in a quantitative evaluation that this new method competes with state of the art algorithms while it is more flexible to incorporate prior knowledge from tracking.