Scene depth estimation from stereo and monocular imagery is critical for extracting 3D information for downstream tasks such as scene understanding. Recently, learning-based methods for depth estimation have received much attention due to their high performance and flexibility in hardware choice. However, collecting ground truth data for supervised training of these algorithms is costly or outright impossible. This circumstance suggests a need for alternative learning approaches that do not require corresponding depth measurements. Indeed, self-supervised learning of depth estimation provides an increasingly popular alternative. It is based on the idea that observed frames can be synthesized from neighboring frames if accurate depth of the scene is known - or in this case, estimated. We show empirically that - contrary to common belief - improvements in image synthesis do not necessitate improvement in depth estimation. Rather, optimizing for image synthesis can result in diverging performance with respect to the main prediction objective - depth. We attribute this diverging phenomenon to aleatoric uncertainties, which originate from data. Based on our experiments on four datasets (spanning street, indoor, and medical) and five architectures (monocular and stereo), we conclude that this diverging phenomenon is independent of the dataset domain and not mitigated by commonly used regularization techniques. To underscore the importance of this finding, we include a survey of methods which use image synthesis, totaling 127 papers over the last six years. This observed divergence has not been previously reported or studied in depth, suggesting room for future improvement of self-supervised approaches which might be impacted the finding.
3D scene understanding from point clouds plays a vital role for various robotic applications. Unfortunately, current state-of-the-art methods use separate neural networks for different tasks like object detection or room layout estimation. Such a scheme has two limitations: 1) Storing and running several networks for different tasks are expensive for typical robotic platforms. 2) The intrinsic structure of separate outputs are ignored and potentially violated. To this end, we propose the first transformer architecture that predicts 3D objects and layouts simultaneously, using point cloud inputs. Unlike existing methods that either estimate layout keypoints or edges, we directly parameterize room layout as a set of quads. As such, the proposed architecture is termed as P(oint)Q(uad)-Transformer. Along with the novel quad representation, we propose a tailored physical constraint loss function that discourages object-layout interference. The quantitative and qualitative evaluations on the public benchmark ScanNet show that the proposed PQ-Transformer succeeds to jointly parse 3D objects and layouts, running at a quasi-real-time (8.91 FPS) rate without efficiency-oriented optimization. Moreover, the new physical constraint loss can improve strong baselines, and the F1-score of the room layout is significantly promoted from 37.9% to 57.9%.
Scene flow estimation is the task to predict the point-wise or pixel-wise 3D displacement vector between two consecutive frames of point clouds or images, which has important application in fields such as service robots and autonomous driving. Although many previous works have explored greatly on scene flow estimation based on point clouds, there are two problems that have not been noticed or well solved before: 1) Points of adjacent frames in repetitive patterns may be wrongly associated due to similar spatial structure in their neighbourhoods; 2) Scene flow between adjacent frames of point clouds with long-distance movement may be inaccurately estimated. To solve the first problem, a novel context-aware set convolution layer is proposed in this paper to exploit contextual structure information of Euclidean space and learn soft aggregation weights for local point features. This design is inspired by human perception of contextual structure information during scene understanding with repetitive patterns. The context-aware set convolution layer is incorporated in a context-aware point feature pyramid module of 3D point clouds for scene flow estimation. For the second problem, an explicit residual flow learning structure is proposed in the residual flow refinement layer to cope with long-distance movement. The experiments and ablation study on FlyingThings3D and KITTI scene flow datasets demonstrate the effectiveness of each proposed component. The qualitative results show that the problems of ambiguous inter-frame association and long-distance movement estimation are well handled. Quantitative results on both FlyingThings3D and KITTI scene flow datasets show that the proposed method achieves state-of-the-art performance, surpassing all other previous works to the best of our knowledge by at least 25%.
Reconstructing 3D object from a single image (RGB or depth) is a fundamental problem in visual scene understanding and yet remains challenging due to its ill-posed nature and complexity in real-world scenes. To address those challenges, we adopt a primitive-based representation for 3D object, and propose a two-stage graph network for primitive-based 3D object estimation, which consists of a sequential proposal module and a graph reasoning module. Given a 2D image, our proposal module first generates a sequence of 3D primitives from input image with local feature attention. Then the graph reasoning module performs joint reasoning on a primitive graph to capture the global shape context for each primitive. Such a framework is capable of taking into account rich geometry and semantic constraints during 3D structure recovery, producing 3D objects with more coherent structure even under challenging viewing conditions. We train the entire graph neural network in a stage-wise strategy and evaluate it on three benchmarks: Pix3D, ModelNet and NYU Depth V2. Extensive experiments show that our approach outperforms the previous state of the arts with a considerable margin.
Automatic building segmentation is an important task for satellite imagery analysis and scene understanding. Most existing segmentation methods focus on the case where the images are taken from directly overhead (i.e., low off-nadir/viewing angle). These methods often fail to provide accurate results on satellite images with larger off-nadir angles due to the higher noise level and lower spatial resolution. In this paper, we propose a method that is able to provide accurate building segmentation for satellite imagery captured from a large range of off-nadir angles. Based on Bayesian deep learning, we explicitly design our method to learn the data noise via aleatoric and epistemic uncertainty modeling. Satellite image metadata (e.g., off-nadir angle and ground sample distance) is also used in our model to further improve the result. We show that with uncertainty modeling and metadata injection, our method achieves better performance than the baseline method, especially for noisy images taken from large off-nadir angles.
Automatically translating images to texts involves image scene understanding and language modeling. In this paper, we propose a novel model, termed RefineCap, that refines the output vocabulary of the language decoder using decoder-guided visual semantics, and implicitly learns the mapping between visual tag words and images. The proposed Visual-Concept Refinement method can allow the generator to attend to semantic details in the image, thereby generating more semantically descriptive captions. Our model achieves superior performance on the MS-COCO dataset in comparison with previous visual-concept based models.
Panoptic scene understanding and tracking of dynamic agents are essential for robots and automated vehicles to navigate in urban environments. As LiDARs provide accurate illumination-independent geometric depictions of the scene, performing these tasks using LiDAR point clouds provides reliable predictions. However, existing datasets lack diversity in the type of urban scenes and have a limited number of dynamic object instances which hinders both learning of these tasks as well as credible benchmarking of the developed methods. In this paper, we introduce the large-scale Panoptic nuScenes benchmark dataset that extends our popular nuScenes dataset with point-wise groundtruth annotations for semantic segmentation, panoptic segmentation, and panoptic tracking tasks. To facilitate comparison, we provide several strong baselines for each of these tasks on our proposed dataset. Moreover, we analyze the drawbacks of the existing metrics for panoptic tracking and propose the novel instance-centric PAT metric that addresses the concerns. We present exhaustive experiments that demonstrate the utility of Panoptic nuScenes compared to existing datasets and make the online evaluation server available at nuScenes.org. We believe that this extension will accelerate the research of novel methods for scene understanding of dynamic urban environments.
Humans can robustly recognize and localize objects by using visual and/or auditory cues. While machines are able to do the same with visual data already, less work has been done with sounds. This work develops an approach for scene understanding purely based on binaural sounds. The considered tasks include predicting the semantic masks of sound-making objects, the motion of sound-making objects, and the depth map of the scene. To this aim, we propose a novel sensor setup and record a new audio-visual dataset of street scenes with eight professional binaural microphones and a 360-degree camera. The co-existence of visual and audio cues is leveraged for supervision transfer. In particular, we employ a cross-modal distillation framework that consists of multiple vision teacher methods and a sound student method -- the student method is trained to generate the same results as the teacher methods do. This way, the auditory system can be trained without using human annotations. To further boost the performance, we propose another novel auxiliary task, coined Spatial Sound Super-Resolution, to increase the directional resolution of sounds. We then formulate the four tasks into one end-to-end trainable multi-tasking network aiming to boost the overall performance. Experimental results show that 1) our method achieves good results for all four tasks, 2) the four tasks are mutually beneficial -- training them together achieves the best performance, 3) the number and orientation of microphones are both important, and 4) features learned from the standard spectrogram and features obtained by the classic signal processing pipeline are complementary for auditory perception tasks. The data and code are released.
Depth sensors used in autonomous driving and gaming systems often report back 3D point clouds. The lack of structure from these sensors does not allow these systems to take advantage of recent advances in convolutional neural networks which are dependent upon traditional filtering and pooling operations. Analogous to image based convolutional architectures, recently introduced graph based architectures afford similar filtering and pooling operations on arbitrary graphs. We adopt these graph based methods to 3D point clouds to introduce a generic vector representation of 3D graphs, we call graph 3D (G3D). We believe we are the first to use large scale transfer learning on 3D point cloud data and demonstrate the discriminant power of our salient latent representation of 3D point clouds on unforeseen test sets. By using our G3D network (G3DNet) as a feature extractor, and then pairing G3D feature vectors with a standard classifier, we achieve the best accuracy on ModelNet10 (93.1%) and ModelNet 40 (91.7%) for a graph network, and comparable performance on the Sydney Urban Objects dataset to other methods. This general-purpose feature extractor can be used as an off-the-shelf component in other 3D scene understanding or object tracking works.
Single-task learning in artificial neural networks will be able to learn the model very well, and the benefits brought by transferring knowledge thus become limited. In this regard, when the number of tasks increases (e.g., semantic segmentation, panoptic segmentation, monocular depth estimation, and 3D point cloud), duplicate information may exist across tasks, and the improvement becomes less significant. Multi-task learning has emerged as a solution to knowledge-transfer issues and is an approach to scene understanding which involves multiple related tasks each with potentially limited training data. Multi-task learning improves generalization by leveraging the domain-specific information contained in the training data of related tasks. In urban management applications such as infrastructure development, traffic monitoring, smart 3D cities, and change detection, automated multi-task data analysis for scene understanding based on the semantic, instance, and panoptic annotation, as well as monocular depth estimation, is required to generate precise urban models. In this study, a common framework for the performance assessment of multi-task learning methods from fixed-wing UAV images for 2D/3D city modeling is presented.
To date, various 3D scene understanding tasks still lack practical and generalizable pre-trained models, primarily due to the intricate nature of 3D scene understanding tasks and their immense variations introduced by camera views, lighting, occlusions, etc. In this paper, we tackle this challenge by introducing a spatio-temporal representation learning (STRL) framework, capable of learning from unlabeled 3D point clouds in a self-supervised fashion. Inspired by how infants learn from visual data in the wild, we explore the rich spatio-temporal cues derived from the 3D data. Specifically, STRL takes two temporally-correlated frames from a 3D point cloud sequence as the input, transforms it with the spatial data augmentation, and learns the invariant representation self-supervisedly. To corroborate the efficacy of STRL, we conduct extensive experiments on three types (synthetic, indoor, and outdoor) of datasets. Experimental results demonstrate that, compared with supervised learning methods, the learned self-supervised representation facilitates various models to attain comparable or even better performances while capable of generalizing pre-trained models to downstream tasks, including 3D shape classification, 3D object detection, and 3D semantic segmentation. Moreover, the spatio-temporal contextual cues embedded in 3D point clouds significantly improve the learned representations.
The scene graph generation (SGG) task aims to detect visual relationship triplets, i.e., subject, predicate, object, in an image, providing a structural vision layout for scene understanding. However, current models are stuck in common predicates, e.g., "on" and "at", rather than informative ones, e.g., "standing on" and "looking at", resulting in the loss of precise information and overall performance. If a model only uses "stone on road" rather than "blocking" to describe an image, it is easy to misunderstand the scene. We argue that this phenomenon is caused by two key imbalances between informative predicates and common ones, i.e., semantic space level imbalance and training sample level imbalance. To tackle this problem, we propose BA-SGG, a simple yet effective SGG framework based on balance adjustment but not the conventional distribution fitting. It integrates two components: Semantic Adjustment (SA) and Balanced Predicate Learning (BPL), respectively for adjusting these imbalances. Benefited from the model-agnostic process, our method is easily applied to the state-of-the-art SGG models and significantly improves the SGG performance. Our method achieves 14.3%, 8.0%, and 6.1% higher Mean Recall (mR) than that of the Transformer model at three scene graph generation sub-tasks on Visual Genome, respectively. Codes are publicly available.
Panorama images have a much larger field-of-view thus naturally encode enriched scene context information compared to standard perspective images, which however is not well exploited in the previous scene understanding methods. In this paper, we propose a novel method for panoramic 3D scene understanding which recovers the 3D room layout and the shape, pose, position, and semantic category for each object from a single full-view panorama image. In order to fully utilize the rich context information, we design a novel graph neural network based context model to predict the relationship among objects and room layout, and a differentiable relationship-based optimization module to optimize object arrangement with well-designed objective functions on-the-fly. Realizing the existing data are either with incomplete ground truth or overly-simplified scene, we present a new synthetic dataset with good diversity in room layout and furniture placement, and realistic image quality for total panoramic 3D scene understanding. Experiments demonstrate that our method outperforms existing methods on panoramic scene understanding in terms of both geometry accuracy and object arrangement. Code is available at https://chengzhag.github.io/publication/dpc.
Specularity prediction is essential to many computer vision applications, giving important visual cues usable in Augmented Reality (AR), Simultaneous Localisation and Mapping (SLAM), 3D reconstruction and material modeling. However, it is a challenging task requiring numerous information from the scene including the camera pose, the geometry of the scene, the light sources and the material properties. Our previous work addressed this task by creating an explicit model using an ellipsoid whose projection fits the specularity image contours for a given camera pose. These ellipsoid-based approaches belong to a family of models called JOint-LIght MAterial Specularity (JOLIMAS), which we have gradually improved by removing assumptions on the scene geometry. However, our most recent approach is still limited to uniformly curved surfaces. This paper generalises JOLIMAS to any surface geometry while improving the quality of specularity prediction, without sacrificing computation performances. The proposed method establishes a link between surface curvature and specularity shape in order to lift the geometric assumptions made in previous work. Contrary to previous work, our new model is built from a physics-based local illumination model namely Torrance-Sparrow, providing an improved reconstruction. Specularity prediction using our new model is tested against the most recent JOLIMAS version on both synthetic and real sequences with objects of various general shapes. Our method outperforms previous approaches in specularity prediction, including the real-time setup, as shown in the supplementary videos.
3D perception of object shapes from RGB image input is fundamental towards semantic scene understanding, grounding image-based perception in our spatially 3-dimensional real-world environments. To achieve a mapping between image views of objects and 3D shapes, we leverage CAD model priors from existing large-scale databases, and propose a novel approach towards constructing a joint embedding space between 2D images and 3D CAD models in a patch-wise fashion -- establishing correspondences between patches of an image view of an object and patches of CAD geometry. This enables part similarity reasoning for retrieving similar CADs to a new image view without exact matches in the database. Our patch embedding provides more robust CAD retrieval for shape estimation in our end-to-end estimation of CAD model shape and pose for detected objects in a single input image. Experiments on in-the-wild, complex imagery from ScanNet show that our approach is more robust than state of the art in real-world scenarios without any exact CAD matches.
Localizing objects and estimating their extent in 3D is an important step towards high-level 3D scene understanding, which has many applications in Augmented Reality and Robotics. We present ODAM, a system for 3D Object Detection, Association, and Mapping using posed RGB videos. The proposed system relies on a deep learning front-end to detect 3D objects from a given RGB frame and associate them to a global object-based map using a graph neural network (GNN). Based on these frame-to-model associations, our back-end optimizes object bounding volumes, represented as super-quadrics, under multi-view geometry constraints and the object scale prior. We validate the proposed system on ScanNet where we show a significant improvement over existing RGB-only methods.
Deep learning has become an integral part of various computer vision systems in recent years due to its outstanding achievements for object recognition, facial recognition, and scene understanding. However, deep neural networks (DNNs) are susceptible to be fooled with nearly high confidence by an adversary. In practice, the vulnerability of deep learning systems against carefully perturbed images, known as adversarial examples, poses a dire security threat in the physical world applications. To address this phenomenon, we present, what to our knowledge, is the first ever image set based adversarial defence approach. Image set classification has shown an exceptional performance for object and face recognition, owing to its intrinsic property of handling appearance variability. We propose a robust deep Bayesian image set classification as a defence framework against a broad range of adversarial attacks. We extensively experiment the performance of the proposed technique with several voting strategies. We further analyse the effects of image size, perturbation magnitude, along with the ratio of perturbed images in each image set. We also evaluate our technique with the recent state-of-the-art defence methods, and single-shot recognition task. The empirical results demonstrate superior performance on CIFAR-10, MNIST, ETH-80, and Tiny ImageNet datasets.
Deep neural networks have proven increasingly important for automotive scene understanding with new algorithms offering constant improvements of the detection performance. However, there is little emphasis on experiences and needs for deployment in embedded environments. We therefore perform a case study of the deployment of two representative object detection networks on an edge AI platform. In particular, we consider RetinaNet for image-based 2D object detection and PointPillars for LiDAR-based 3D object detection. We describe the modifications necessary to convert the algorithms from a PyTorch training environment to the deployment environment taking into account the available tools. We evaluate the runtime of the deployed DNN using two different libraries, TensorRT and TorchScript. In our experiments, we observe slight advantages of TensorRT for convolutional layers and TorchScript for fully connected layers. We also study the trade-off between runtime and performance, when selecting an optimized setup for deployment, and observe that quantization significantly reduces the runtime while having only little impact on the detection performance.
Depth estimation, as a necessary clue to convert 2D images into the 3D space, has been applied in many machine vision areas. However, to achieve an entire surrounding 360-degree geometric sensing, traditional stereo matching algorithms for depth estimation are limited due to large noise, low accuracy, and strict requirements for multi-camera calibration. In this work, for a unified surrounding perception, we introduce panoramic images to obtain larger field of view. We extend PADENet first appeared in our previous conference work for outdoor scene understanding, to perform panoramic monocular depth estimation with a focus for indoor scenes. At the same time, we improve the training process of the neural network adapted to the characteristics of panoramic images. In addition, we fuse traditional stereo matching algorithm with deep learning methods and further improve the accuracy of depth predictions. With a comprehensive variety of experiments, this research demonstrates the effectiveness of our schemes aiming for indoor scene perception.
In this paper, we present a system for incrementally reconstructing a dense 3D model of the geometry of an outdoor environment using a single monocular camera attached to a moving vehicle. Dense models provide a rich representation of the environment facilitating higher-level scene understanding, perception, and planning. Our system employs dense depth prediction with a hybrid mapping architecture combining state-of-the-art sparse features and dense fusion-based visual SLAM algorithms within an integrated framework. Our novel contributions include design of hybrid sparse-dense camera tracking and loop closure, and scale estimation improvements in dense depth prediction. We use the motion estimates from the sparse method to overcome the large and variable inter-frame displacement typical of outdoor vehicle scenarios. Our system then registers the live image with the dense model using whole-image alignment. This enables the fusion of the live frame and dense depth prediction into the model. Global consistency and alignment between the sparse and dense models are achieved by applying pose constraints from the sparse method directly within the deformation of the dense model. We provide qualitative and quantitative results for both trajectory estimation and surface reconstruction accuracy, demonstrating competitive performance on the KITTI dataset. Qualitative results of the proposed approach are illustrated in https://youtu.be/Pn2uaVqjskY. Source code for the project is publicly available at the following repository https://github.com/robotvisionmu/DenseMonoSLAM.