Depth Privileged Object Detection in Indoor Scenes via Deformation Hallucination

A multi-cue guidance network for depth completion

In this paper we study the problem of object detection for RGB-D images using semantically rich image and depth features. We propose a new geocentric embedding for depth images that encodes height above ground and angle with gravity for each pixel in addition to the horizontal disparity. We demonstrate that this geocentric embedding works better than using raw depth images for learning feature representations with convolutional neural networks. Our final object detection system achieves an average precision of 37.3%, which is a 56% relative improvement over existing methods. We then focus on the task of instance segmentation where we label pixels belonging to object instances found by our detector. For this task, we propose a decision forest approach that classifies pixels in the detection window as foreground or background using a family of unary and binary tests that query shape and geocentric pose features. Finally, we use the output from our object detectors in an existing superpixel classification framework for semantic scene segmentation and achieve a 24% relative improvement over current state-of-the-art for the object categories that we study. We believe advances such as those represented in this paper will facilitate the use of perception in fields like robotics.KeywordsRGB-D perceptionobject detectionobject segmentation

RGB-D-based object recognition has been enthusiastically investigated in the past few years. RGB and depth images provide useful and complementary information. Fusing RGB and depth features can significantly increase the accuracy of object recognition. However, previous works just simply take the depth image as the fourth channel of the RGB image and concatenate the RGB and depth features, ignoring the different power of RGB and depth information for different objects. In this paper, a new method which contains three different classifiers is proposed to fuse features extracted from RGB image and depth image for RGB-D-based object recognition. Firstly, a RGB classifier and a depth classifier are trained by cross-validation to get the accuracy difference between RGB and depth features for each object. Then a variant RGB-D classifier is trained with different initialization parameters for each class according to the accuracy difference. The variant RGB-D-classifier can result in a more robust classification performance. The proposed method is evaluated on two benchmark RGB-D datasets. Compared with previous methods, ours achieves comparable performance with the state-of-the-art method.

There has been significant progress in recognition of outdoor scenes but indoor scene recognition is still an challenge. This is due to the high appearance fluctuation of indoor situations. With the recent developments in indoor and mobile robotics, identifying the indoor scenes has gained importance. Many approaches have been proposed to detect scenes using object detection and geotags. In contrast, the proposal of this paper uses the convolutional neural network which has gained importance with advancement in machine learning methodologies. Our method has higher efficiency than the existing models as we try to classify the environment as a whole rather than using object identification for the same. We test this approach on our dataset which consists of RGB and also depth images of common locations present in academic environments such as class rooms, labs etc. The proposed approach performs better than previous ones with accuracy up to 98%.

In this paper, we address the problems of contour detection, bottom-up grouping, object detection and semantic segmentation on RGB-D data. We focus on the challenging setting of cluttered indoor scenes, and evaluate our approach on the recently introduced NYU-Depth V2 (NYUD2) dataset (Silberman et al., ECCV, 2012). We propose algorithms for object boundary detection and hierarchical segmentation that generalize the $$gPb-ucm$$gPb-ucm approach of Arbelaez et al. (TPAMI, 2011) by making effective use of depth information. We show that our system can label each contour with its type (depth, normal or albedo). We also propose a generic method for long-range amodal completion of surfaces and show its effectiveness in grouping. We train RGB-D object detectors by analyzing and computing histogram of oriented gradients on the depth image and using them with deformable part models (Felzenszwalb et al., TPAMI, 2010). We observe that this simple strategy for training object detectors significantly outperforms more complicated models in the literature. We then turn to the problem of semantic segmentation for which we propose an approach that classifies superpixels into the dominant object categories in the NYUD2 dataset. We design generic and class-specific features to encode the appearance and geometry of objects. We also show that additional features computed from RGB-D object detectors and scene classifiers further improves semantic segmentation accuracy. In all of these tasks, we report significant improvements over the state-of-the-art.

Many perception problems in robotics such as object recognition, scene understanding, and mapping are tackled using scale-invariant interest points extracted from intensity images. Since interest points describe only local portions of objects and scenes, they offer robustness to clutter, occlusions, and intra-class variation. In this paper, we present an efficient approximate algorithm to extract surface normal interest points (SNIPs) in corners and blob-like surface regions from depth images. The interest points are detected on characteristic scales that indicate their spatial extent. Our method is able to cope with irregularly sampled, noisy measurements which are typical to depth imaging devices. It also offers a trade-off between computational speed and accuracy which allows our approach to be applicable in a wide range of problem sets. We evaluate our approach on depth images of basic geometric shapes, more complex objects, and indoor scenes.

Bericht des Nationalen Referenzzentrums (NRZ) für gramnegative Krankenhauserreger Zeitraum 1. Januar 2016 – 31. Dezember 2016

Recognizing objects and estimating their poses have a wide range of application in robotics. For instance, to grasp objects, robots need the position and orientation of objects in 3D. The task becomes challenging in a cluttered environment with different types of objects. A popular approach to tackle this problem is to utilize a deep neural network for object recognition. However, deep learning-based object detection in cluttered environments requires a substantial amount of data. Collection of these data requires time and extensive human labor for manual labeling. In this study, our objective was the development and validation of a deep object recognition framework using a synthetic depth image dataset. We synthetically generated a depth image dataset of 22 objects randomly placed in a 0.5 m × 0.5 m × 0.1 m box, and automatically labeled all objects with an occlusion rate below 70%. Faster Region Convolutional Neural Network (R-CNN) architecture was adopted for training using a dataset of 800,000 synthetic depth images, and its performance was tested on a real-world depth image dataset consisting of 2000 samples. Deep object recognizer has 40.96% detection accuracy on the real depth images and 93.5% on the synthetic depth images. Training the deep learning model with noise-added synthetic images improves the recognition accuracy for real images to 46.3%. The object detection framework can be trained on synthetically generated depth data, and then employed for object recognition on the real depth data in a cluttered environment. Synthetic depth data-based deep object detection has the potential to substantially decrease the time and human effort required for the extensive data collection and labeling.

Depth cameras started to gain popularity in agricultural applications during the last years. This type of cameras has been implemented mainly for three-dimensional (3D) reconstruction of objects in indoor or outdoor sceneries. The use of such cameras in the construction of 3D models for simulation purposes of complex structures that are usually met in nature, such as trees and other plants, is a great challenge. Remarkably, agricultural environments are extremely complex. Thus, the proper setup and implementation of such technologies is particularly important in order to attain useable data. So far, the depth information collected using these cameras varies among different objects’ structure and sensing conditions due to the uncertainty of the outdoor environment. The use of a specific methodology using color and depth images gives the opportunity to extract geometrical characteristics information about point clouds of the targeted objects. This chapter explores the different technologies used by depth cameras and presents several applications concerning indoor and outdoor environments by presenting indicative scenarios for agricultural applications. Towards that direction, a 3D reconstruction of trees was established producing point clouds from Red Green Blue Depth (RGB-D) images acquired in real field conditions. The point cloud samples of trees were collected using an unmanned ground vehicle (UGV) and imported in Gazebo in order to visualize a simulation of the environment. This simulation technique can be used for testing and evaluating the navigation of robotic systems. By further analyzing the resulted 3D point clouds, various geometrical measurements of the simulated samples, such as the volume or the height of tree canopies, can be calculated. Possible weaknesses of this procedure are mainly attributed to the camera’s limitations and the sampling parameters. However, results show that it is possible to establish a suitable simulation environment to implement it in several agricultural applications by utilizing automated unmanned robotic platforms.KeywordsDepth camerasUnmanned ground vehiclesScene reconstruction3D point cloudAgricultural robotsSituation awareness

We propose a novel approach for multi-view object detection in 3D scenes reconstructed from RGB-D sensor. We utilize shape based representation using local shape context descriptors along with the voting strategy which is supported by unsupervised object proposals generated from 3D point cloud data. Our algorithm starts with a single-view object detection where object proposals generated in 3D space are combined with object specific hypotheses generated by the voting strategy. To tackle the multi-view setting, the data association between multiple views enabled view registration and 3D object proposals. The evidence from multiple views is combined in simple bayesian setting. The approach is evaluated on the Washington RGB-D scenes datasets [1], [2] containing several classes of objects in a table top setting. We evaluated our approach against the other state-of-the-art methods and demonstrated superior performance on the same dataset.

RGB-D salient object detection aims at identifying the most attractive parts from a RGB image and its corresponding depth image, which has been widely applied in many computer vision tasks. However, there are still two challenges: (1) how to quickly and effectively integrate the cross-modal features from the RGB-D data; and (2) how to mitigate the negative impact from the low-quality depth map. The previous methods mostly employ a two-stream architecture which adopts two backbone network to process RGB-D data and ignore the quality of depth map. In this paper, we propose a guided residual network to address these two issues. On the one hand, we design a simpler and efficient depth branch only using one convolutional layer and three residual modules to extract depth features instead of employing a pre-trained backbone to handle the depth data, and fuse RGB features and depth features in a multi-scale manner for refinement with top-down guidance. On the other hand, we add adaptive weight to depth maps to control the fusion between them, which mitigates the negative influence of unreliable depth map. Experimental results compared with 13 state-of-the-art methods on 7 datasets demonstrate the validity of the proposed approach both quantitatively and qualitatively, especially in efficiency (102 FPS) and compactness (64.2 MB).

Multi-scale object detection in natural scenes is still challenging. To enhance the multi-scale perception capability, some algorithms combine the lower-level and higher-level information via multi-scale feature fusion strategies. However, the inherent spatial properties among instances and relations between foreground and background are ignored. In addition, the human-defined “center-based” regression quality evaluation strategy, predicting a high-to-low score based on a linear relationship with the distance to the center of ground-truth box, is not robust to scale-variant objects. In this work, we propose a Depth-Guided Progressive Network (DGPNet) for multi-scale object detection. Specifically, besides the prediction of classification and localization, the depth is estimated and used to guide the image features in a weighted manner to obtain a better spatial representation. Therefore, depth estimation and 2D object detection are simultaneously learned via a unified network, where the depth features are merged as auxiliary information into the detection branch to enhance the discrimination among multi-scale objects. Moreover, to overcome the difficulty of empirically fitting the localization quality function, high-quality predicted boxes on scale-variant objects are more adaptively obtained by an IoU-aware progressive sampling strategy. We divide the sampling process into two stages, i.e., “statistical-aware” and “IoU-aware”. The former selects thresholds for positive samples based on statistical characteristics of multi-scale instances, and the latter further selects high-quality samples by IoU on the basis of the former. Therefore, the final ranking scores better reflect the quality of localization. Experiments verify that our method outperforms state-of-the-art methods on the KINS and Cityscapes dataset.

This paper proposes recognition and grasping objects from 3D environment by combining depth and color stereo image in the mobile picking robot system. To do this task, the followings are done. Firstly, an image processing system including Kinect camera sensor is described. Secondly, RGB color map and new depth map for image inpainting are obtained using Kinect SDK mapping function to align RGB image with depth image. Thirdly, the new depth map are segmented to distinguish between the image background and the objects that should be recognized. The feature colours are generated based on colour histograms. Euclidean distance is used to measure the similarity between the feature vectors computed from the colour image and the feature vectors stored in a database. Fourthly, by converting RGB map and new depth map into 3D point clouds, an algorithm for localizing handle-like grasp affordances is proposed. The main idea is to search the point cloud for neighborhoods that satisfy handle-like grasp affordances and can be grasped by the end-effector of the manipulator. Finally, the effectiveness of the proposed algorithms is verified by using experiment. The experimental results show that the mobile picking robot successfully detects an object and finds its grasping points with an acceptable small error.

In order to obtain better super-resolution reconstruction results of depth images, this paper constructs a multi-scale color image guidance depth image super-resolution reconstruction convolutional neural network. In this paper, the multi-scale fusion method is used to realize the guidance of high resolution (HR) color image features to low resolution (LR) depth image features, which is beneficial to the restoration of image details. In the process of extracting features from LR depth images, a multiple receptive field residual block (MRFRB) is constructed to extract and fuse the features of different receptive fields, and then connect and fuse the features of each MRFRB output to obtain global fusion features. Finally, the HR depth image is obtained through sub-pixel convolution layer and global fusion features. The experimental results show that the super-resolution image obtained by this method alleviates the edge distortion and artifact problems, and has better visual effects.

Human action recognition and localization is a challenging vision task with promising applications. To tackle this problem, recently developed commodity depth sensor (e.g., Microsoft Kinect) has opened up new opportunities with several developed human motion features based on depth image for action representation. However, how depth information can be effectively adopted in the middle or high level representation in action detection, in particular, the depth induced three dimensional contextual information for modeling interactions between human-human, human-object and human-surroundings has yet been explored. In this paper, we propose a novel action recognition and localization framework which effectively fuses depth-induced contextual information from different levels of the processing pipeline for understanding various interactions. First, depth image is combined with grayscale image for more robust human subject and object detection. Second, three dimensional spatial and temporal relationship among human subjects or objects is represented based on the combination of grayscale and depth images. Third, depth information is further utilized to represent different types of indoor scenes. Finally, we fuse these multiple stage depth-induced contextual information to yield an unified action detection framework. Extensive experiments on a challenging grayscale + depth human action detection benchmark database demonstrate the effectiveness of the depth-induced contextual information and the high detection accuracy of the proposed framework.