Video object segmentation (VOS) describes the task of segmenting a set of objects in each frame of a video. In the semi-supervised setting, the first mask of each object is provided at test time. Following the one-shot principle, fine-tuning VOS methods train a segmentation model separately on each given object mask. However, recently the VOS community has deemed such a test time optimization and its impact on the test runtime as unfeasible. To mitigate the inefficiencies of previous fine-tuning approaches, we present efficient One-Shot Video Object Segmentation (e-OSVOS). In contrast to most VOS approaches, e-OSVOS decouples the object detection task and predicts only local segmentation masks by applying a modified version of Mask R-CNN. The one-shot test runtime and performance are optimized without a laborious and handcrafted hyperparameter search. To this end, we meta learn the model initialization and learning rates for the test time optimization. To achieve optimal learning behavior, we predict individual learning rates at a neuron level. Furthermore, we apply an online adaptation to address the common performance degradation throughout a sequence by continuously fine-tuning the model on previous mask predictions supported by a frame-to-frame bounding box propagation. e-OSVOS provides state-of-the-art results on DAVIS 2016, DAVIS 2017, and YouTube-VOS for one-shot fine-tuning methods while reducing the test runtime substantially. Code is available at https://github.com/dvl-tum/e-osvos.
This paper presents the first active object mapping framework for complex robotic manipulation and autonomous perception tasks. The framework is built on an object SLAM system integrated with a simultaneous multi-object pose estimation process that is optimized for robotic grasping. Aiming to reduce the observation uncertainty on target objects and increase their pose estimation accuracy, we also design an object-driven exploration strategy to guide the object mapping process, enabling autonomous mapping and high-level perception. Combining the mapping module and the exploration strategy, an accurate object map that is compatible with robotic grasping can be generated. Additionally, quantitative evaluations also indicate that the proposed framework has a very high mapping accuracy. Experiments with manipulation (including object grasping and placement) and augmented reality significantly demonstrate the effectiveness and advantages of our proposed framework.
To determine if a skill can be executed in any given environment, a robot needs to learn the preconditions for the skill. As robots begin to operate in dynamic and unstructured environments, precondition models will need to generalize to variable number of objects with different shapes and sizes. In this work, we focus on learning precondition models for manipulation skills in unconstrained environments. Our work is motivated by the intuition that many complex manipulation tasks, with multiple objects, can be simplified by focusing on less complex pairwise object relations. We propose an object-relation model that learns continuous representations for these pairwise object relations. Our object-relation model is trained completely in simulation, and once learned, is used by a separate precondition model to predict skill preconditions for real world tasks. We evaluate our precondition model on $3$ different manipulation tasks: sweeping, cutting, and unstacking. We show that our approach leads to significant improvements in predicting preconditions for all 3 tasks, across objects of different shapes and sizes.
FPN is a common component used in object detectors, it supplements multi-scale information by adjacent level features interpolation and summation. However, due to the existence of nonlinear operations and the convolutional layers with different output dimensions, the relationship between different levels is much more complex, the pixel-wise summation is not an efficient approach. In this paper, we first analyze the design defects from pixel level and feature map level. Then, we design a novel parameter-free feature pyramid networks named Dual Refinement Feature Pyramid Networks (DRFPN) for the problems. Specifically, DRFPN consists of two modules: Spatial Refinement Block (SRB) and Channel Refinement Block (CRB). SRB learns the location and content of sampling points based on contextual information between adjacent levels. CRB learns an adaptive channel merging method based on attention mechanism. Our proposed DRFPN can be easily plugged into existing FPN-based models. Without bells and whistles, for two-stage detectors, our model outperforms different FPN-based counterparts by 1.6 to 2.2 AP on the COCO detection benchmark, and 1.5 to 1.9 AP on the COCO segmentation benchmark. For one-stage detectors, DRFPN improves anchor-based RetinaNet by 1.9 AP and anchor-free FCOS by 1.3 AP when using ResNet50 as backbone. Extensive experiments verifies the robustness and generalization ability of DRFPN. The code will be made publicly available.
Capsule Networks, as alternatives to Convolutional Neural Networks, have been proposed to recognize objects from images. The current literature demonstrates many advantages of CapsNets over CNNs. However, how to create explanations for individual classifications of CapsNets has not been well explored. The widely used saliency methods are mainly proposed for explaining CNN-based classifications; they create saliency map explanations by combining activation values and the corresponding gradients, e.g., Grad-CAM. These saliency methods require a specific architecture of the underlying classifiers and cannot be trivially applied to CapsNets due to the iterative routing mechanism therein. To overcome the lack of interpretability, we can either propose new post-hoc interpretation methods for CapsNets or modifying the model to have build-in explanations. In this work, we explore the latter. Specifically, we propose interpretable Graph Capsule Networks (GraCapsNets), where we replace the routing part with a multi-head attention-based Graph Pooling approach. In the proposed model, individual classification explanations can be created effectively and efficiently. Our model also demonstrates some unexpected benefits, even though it replaces the fundamental part of CapsNets. Our GraCapsNets achieve better classification performance with fewer parameters and better adversarial robustness, when compared to CapsNets. Besides, GraCapsNets also keep other advantages of CapsNets, namely, disentangled representations and affine transformation robustness.
Object detection is a comprehensively studied problem in autonomous driving. However, it has been relatively less explored in the case of fisheye cameras. The standard bounding box fails in fisheye cameras due to the strong radial distortion, particularly in the image's periphery. We explore better representations like oriented bounding box, ellipse, and generic polygon for object detection in fisheye images in this work. We use the IoU metric to compare these representations using accurate instance segmentation ground truth. We design a novel curved bounding box model that has optimal properties for fisheye distortion models. We also design a curvature adaptive perimeter sampling method for obtaining polygon vertices, improving relative mAP score by 4.9% compared to uniform sampling. Overall, the proposed polygon model improves mIoU relative accuracy by 40.3%. It is the first detailed study on object detection on fisheye cameras for autonomous driving scenarios to the best of our knowledge. The dataset comprising of 10,000 images along with all the object representations ground truth will be made public to encourage further research. We summarize our work in a short video with qualitative results at https://youtu.be/iLkOzvJpL-A.
Analyzing complex scenes with Deep Neural Networks is a challenging task, particularly when images contain multiple objects that partially occlude each other. Existing approaches to image analysis mostly process objects independently and do not take into account the relative occlusion of nearby objects. In this paper, we propose a deep network for multi-object instance segmentation that is robust to occlusion and can be trained from bounding box supervision only. Our work builds on Compositional Networks, which learn a generative model of neural feature activations to locate occluders and to classify objects based on their non-occluded parts. We extend their generative model to include multiple objects and introduce a framework for efficient inference in challenging occlusion scenarios. In particular, we obtain feed-forward predictions of the object classes and their instance and occluder segmentations. We introduce an Occlusion Reasoning Module (ORM) that locates erroneous segmentations and estimates the occlusion order to correct them. The improved segmentation masks are, in turn, integrated into the network in a top-down manner to improve the image classification. Our experiments on the KITTI INStance dataset (KINS) and a synthetic occlusion dataset demonstrate the effectiveness and robustness of our model at multi-object instance segmentation under occlusion. Code is publically available at https://github.com/XD7479/Multi-Object-Occlusion.
Recent advances in 3D semantic scene understanding have shown impressive progress in 3D instance segmentation, enabling object-level reasoning about 3D scenes; however, a finer-grained understanding is required to enable interactions with objects and their functional understanding. Thus, we propose the task of part-based scene understanding of real-world 3D environments: from an RGB-D scan of a scene, we detect objects, and for each object predict its decomposition into geometric part masks, which composed together form the complete geometry of the observed object. We leverage an intermediary part graph representation to enable robust completion as well as building of part priors, which we use to construct the final part mask predictions. Our experiments demonstrate that guiding part understanding through part graph to part prior-based predictions significantly outperforms alternative approaches to the task of semantic part completion.
The 3D world limits the human body pose and the human body pose conveys information about the surrounding objects. Indeed, from a single image of a person placed in an indoor scene, we as humans are adept at resolving ambiguities of the human pose and room layout through our knowledge of the physical laws and prior perception of the plausible object and human poses. However, few computer vision models fully leverage this fact. In this work, we propose an end-to-end trainable model that perceives the 3D scene from a single RGB image, estimates the camera pose and the room layout, and reconstructs both human body and object meshes. By imposing a set of comprehensive and sophisticated losses on all aspects of the estimations, we show that our model outperforms existing human body mesh methods and indoor scene reconstruction methods. To the best of our knowledge, this is the first model that outputs both object and human predictions at the mesh level, and performs joint optimization on the scene and human poses.
We introduce Neural Deformation Graphs for globally-consistent deformation tracking and 3D reconstruction of non-rigid objects. Specifically, we implicitly model a deformation graph via a deep neural network. This neural deformation graph does not rely on any object-specific structure and, thus, can be applied to general non-rigid deformation tracking. Our method globally optimizes this neural graph on a given sequence of depth camera observations of a non-rigidly moving object. Based on explicit viewpoint consistency as well as inter-frame graph and surface consistency constraints, the underlying network is trained in a self-supervised fashion. We additionally optimize for the geometry of the object with an implicit deformable multi-MLP shape representation. Our approach does not assume sequential input data, thus enabling robust tracking of fast motions or even temporally disconnected recordings. Our experiments demonstrate that our Neural Deformation Graphs outperform state-of-the-art non-rigid reconstruction approaches both qualitatively and quantitatively, with 64% improved reconstruction and 62% improved deformation tracking performance.
Humans can covertly track the position of an object, even if the object is temporarily occluded. What are the neural mechanisms underlying our capacity to track moving objects when there is no physical stimulus for the brain to track? One possibility is that the brain 'fills-in' information about imagined objects using internally generated representations similar to those generated by feed-forward perceptual mechanisms. Alternatively, the brain might deploy a higher order mechanism, for example using an object tracking model that integrates visual signals and motion dynamics. In the present study, we used EEG and time-resolved multivariate pattern analyses to investigate the spatial processing of visible and imagined objects. Participants tracked an object that moved in discrete steps around fixation, occupying six consecutive locations. They were asked to imagine that the object continued on the same trajectory after it disappeared and move their attention to the corresponding positions. Time-resolved decoding of EEG data revealed that the location of the visible stimuli could be decoded shortly after image onset, consistent with early retinotopic visual processes. For processing of unseen/imagined positions, the patterns of neural activity resembled stimulus-driven mid-level visual processes, but were detected earlier than perceptual mechanisms, implicating an anticipatory and more variable tracking mechanism. Encoding models revealed that spatial representations were much weaker for imagined than visible stimuli. Monitoring the position of imagined objects thus utilises similar perceptual and attentional processes as monitoring objects that are actually present, but with different temporal dynamics. These results indicate that internally generated representations rely on top-down processes, and their timing is influenced by the predictability of the stimulus.
A key challenge in robot teaching is grasp-type recognition with a single RGB image and a target object name. Here, we propose a simple yet effective pipeline to enhance learning-based recognition by leveraging a prior distribution of grasp types for each object. In the pipeline, a convolutional neural network (CNN) recognizes the grasp type from an RGB image. The recognition result is further corrected using the prior distribution (i.e., affordance), which is associated with the target object name. Experimental results showed that the proposed method outperforms both a CNN-only and an affordance-only method. The results highlight the effectiveness of linguistically-driven object affordance for enhancing grasp-type recognition in robot teaching.
Medical image slice interpolation is an active field of research. The methods for this task can be categorized into two broad groups: intensity-based and object-based interpolation methods. While intensity-based methods are generally easier to perform and less computationally expensive, object-based methods are capable of producing more accurate results and account for deformable changes in the objects within the slices. In this paper, performance of two well-known object-based interpolation methods is analyzed and compared. Here, a deformable registration-based method specifically designed for medical applications and a learning-based method, trained for video frame interpolation, are considered. While the deformable registration-based technique is capable of accurate modeling of the changes in the shapes of the objects within slices, the learning-based method is able to produce results with similar accuracy, but with a much sharper appearance in a fraction of the time. This is despite the fact that the learning-based approach is not trained on medical images and rather is trained using regular video footage. However, experiments show that the method is capable of accurate slice interpolation results.
Here we introduce three-dimensional single-shot ptychography (3DSSP). 3DSSP leverages an additional constraint unique to the single-shot geometry to deconvolve multiple 2D planes of a 3D object. Numeric simulations and analytic calculations demonstrate that 3DSSP reconstructs multiple planes in an extended 3D object with a minimum separation consistent with the depth of field for a conventional microscope. We experimentally demonstrate 3DSSP by reconstructing orthogonal hair strands axially separated by 5 mm. Three-dimensional single-shot ptychography provides a pathway towards volumetric imaging of dynamically evolving systems on ultrafast timescales.
When studying how people search for objects in scenes, the inhomogeneity of the visual field is often ignored. Due to physiological limitations peripheral vision is blurred and mainly uses coarse-grained information (i.e., low spatial frequencies) for selecting saccade targets, whereas high-acuity central vision uses fine-grained information (i.e., high spatial frequencies) for analysis of details. Here we investigated how spatial frequencies and color affect object search in real-world scenes. Using gaze-contingent filters we attenuated high or low frequencies in central or peripheral vision while viewers searched color or grayscale scenes. Results showed that peripheral filters and central high-pass filters hardly affected search accuracy, whereas accuracy dropped drastically with central low-pass filters. Peripheral filtering increased the time to localize the target by decreasing saccade amplitudes and increasing number and duration of fixations. The use of coarse-grained information in the periphery was limited to color scenes. Central filtering increased the time to verify target identity instead, especially with low-pass filters. We conclude that peripheral vision is critical for object localization and central vision is critical for object identification. Visual guidance during peripheral object localization is dominated by low-frequency color information, whereas high-frequency information, relatively independent of color, is most important for object identification in central vision.
In this study, we investigate the problem of tracking objects with unknown shapes using three-dimensional (3D) point cloud data. We propose a Gaussian process-based model to jointly estimate object kinematics, including position, orientation and velocities, together with the shape of the object for online and offline applications. We describe the unknown shape by a radial function in 3D, and induce a correlation structure via a Gaussian process. Furthermore, we propose an efficient algorithm to reduce the computational complexity of working with 3D data. This is accomplished by casting the tracking problem into projection planes which are attached to the object's local frame. The resulting algorithms can process 3D point cloud data and accomplish tracking of a dynamic object. Furthermore, they provide analytical expressions for the representation of the object shape in 3D, together with confidence intervals. The confidence intervals, which quantify the uncertainty in the shape estimate, can later be used for solving the gating and association problems inherent in object tracking. The performance of the methods is demonstrated both on simulated and real data. The results are compared with an existing random matrix model, which is commonly used for extended object tracking in the literature.
Computed Tomography is a powerful imaging technique that allows non-destructive visualization of the interior of physical objects in different scientific areas. In traditional reconstruction techniques the object of interest is mostly considered to be static, which gives artefacts if the object is moving during the data acquisition. In this paper we present a method that, given only scan results of multiple successive scans, can estimate the motion and correct the CT-images for this motion assuming that the motion field is smooth over the complete domain using optical flow. The proposed method is validated on simulated scan data. The main contribution is that we show we can use the optical flow technique from imaging to correct CT-scan images for motion.
We report a compact, cost-effective and field-portable lensless imaging platform for quantitative microscopy. In this platform, the object is placed on top of an image sensor chip without using any lens. We use a low-cost galvo scanner to rapidly scan an unknown laser speckle pattern on the object. To address the positioning repeatability and accuracy issues, we directly recover the positional shifts of the speckle pattern based on the phase correlation of the captured images. To bypass the resolution limit set by the imager pixel size, we employ a sub-sampled ptychographic phase retrieval process to recover the complex object. We validate our approach using a resolution target, a phase target, and a biological sample. Our results show that accurate, high-quality complex images can be obtained from a lensless dataset with as few as ~10 images. We also demonstrate the reported approach to achieve a 6.4 mm by 4.6 mm field of view and a half pitch resolution of 1 miron. The reported approach may provide a quantitative lensless imaging strategy for addressing point-of-care, global-health, and telemedicine related challenges.
In this paper, we propose an advanced framework of ghost edge imaging, named compressed ghost edge imaging (CGEI). In the scheme, a set of structured speckle patterns with pixel shifting are illuminated on an unknown object, and the output is collected by a bucket detector without any spatial resolution. By using compressed sensing algorithm, we obtain the horizontal and vertical edge information of the unknown object with the bucket detector detection results and the known structured speckle patterns. The edge is finally constructed by the two-dimentional edge information. The experimental and numerical simulations results show that the proposed scheme has a higher quality and reduces the number of measurements, in comparison with the existed edge detection schemes based on ghost imaging.
We introduced a new kind of patterns named Special-Hadamard patterns, which could be used as structured illuminations of computational ghost imaging. Special-Hadamard patterns can get a better image quality than Hadamard patterns in a noisy environment. We can completely reconstruct the original object in a noiseless environment by using Special-Hadamard patterns, and the size of object also can be adjusted arbitrarily, these advantages cannot be achieved by other common patterns. We also performed simulations to compare the results of Special Hadamard patterns with the results of Hadamard patterns. We found Special Hadamard patterns can greatly improve the image quality of computational ghost imaging.