Recently, the vision transformer has achieved great success by pushing the state-of-the-art of various vision tasks. One of the most challenging problems in the vision transformer is that the large sequence length of image tokens leads to high computational cost (quadratic complexity). A popular solution to this problem is to use a single pooling operation to reduce the sequence length. This paper considers how to improve existing vision transformers, where the pooled feature extracted by a single pooling operation seems less powerful. To this end, we note that pyramid pooling has been demonstrated to be effective in various vision tasks owing to its powerful ability in context abstraction. However, pyramid pooling has not been explored in backbone network design. To bridge this gap, we propose to adapt pyramid pooling to Multi-Head Self-Attention (MHSA) in the vision transformer, simultaneously reducing the sequence length and capturing powerful contextual features. Plugged with our pooling-based MHSA, we build a universal vision transformer backbone, dubbed Pyramid Pooling Transformer (P2T). Extensive experiments demonstrate that, when applied P2T as the backbone network, it shows substantial superiority in various vision tasks such as image classification, semantic segmentation, object detection, and instance segmentation, compared to previous CNN- and transformer-based networks. The code will be released at https://github.com/yuhuan-wu/P2T.
In this work, we introduce the new scene understanding task of Part-aware Panoptic Segmentation (PPS), which aims to understand a scene at multiple levels of abstraction, and unifies the tasks of scene parsing and part parsing. For this novel task, we provide consistent annotations on two commonly used datasets: Cityscapes and Pascal VOC. Moreover, we present a single metric to evaluate PPS, called Part-aware Panoptic Quality (PartPQ). For this new task, using the metric and annotations, we set multiple baselines by merging results of existing state-of-the-art methods for panoptic segmentation and part segmentation. Finally, we conduct several experiments that evaluate the importance of the different levels of abstraction in this single task.
This paper tackles the high computational/space complexity associated with Multi-Head Self-Attention (MHSA) in vanilla vision transformers. To this end, we propose Hierarchical MHSA (H-MHSA), a novel approach that computes self-attention in a hierarchical fashion. Specifically, we first divide the input image into patches as commonly done, and each patch is viewed as a token. Then, the proposed H-MHSA learns token relationships within local patches, serving as local relationship modeling. Then, the small patches are merged into larger ones, and H-MHSA models the global dependencies for the small number of the merged tokens. At last, the local and global attentive features are aggregated to obtain features with powerful representation capacity. Since we only calculate attention for a limited number of tokens at each step, the computational load is reduced dramatically. Hence, H-MHSA can efficiently model global relationships among tokens without sacrificing fine-grained information. With the H-MHSA module incorporated, we build a family of Hierarchical-Attention-based Transformer Networks, namely HAT-Net. To demonstrate the superiority of HAT-Net in scene understanding, we conduct extensive experiments on fundamental vision tasks, including image classification, semantic segmentation, object detection, and instance segmentation. Therefore, HAT-Net provides a new perspective for vision transformers. Code and pretrained models are available at https://github.com/yun-liu/HAT-Net.
Inferring representations of 3D scenes from 2D observations is a fundamental problem of computer graphics, computer vision, and artificial intelligence. Emerging 3D-structured neural scene representations are a promising approach to 3D scene understanding. In this work, we propose a novel neural scene representation, Light Field Networks or LFNs, which represent both geometry and appearance of the underlying 3D scene in a 360-degree, four-dimensional light field parameterized via a neural implicit representation. Rendering a ray from an LFN requires only a single network evaluation, as opposed to hundreds of evaluations per ray for ray-marching or volumetric based renderers in 3D-structured neural scene representations. In the setting of simple scenes, we leverage meta-learning to learn a prior over LFNs that enables multi-view consistent light field reconstruction from as little as a single image observation. This results in dramatic reductions in time and memory complexity, and enables real-time rendering. The cost of storing a 360-degree light field via an LFN is two orders of magnitude lower than conventional methods such as the Lumigraph. Utilizing the analytical differentiability of neural implicit representations and a novel parameterization of light space, we further demonstrate the extraction of sparse depth maps from LFNs.
Autonomous robotics is critically affected by the robustness of its scene understanding algorithms. We propose a two-axis pipeline based on polarization indices to analyze dynamic urban scenes. As robots evolve in unknown environments, they are prone to encountering specular obstacles. Usually, specular phenomena are rarely taken into account by algorithms which causes misinterpretations and erroneous estimates. By exploiting all the light properties, systems can greatly increase their robustness to events. In addition to the conventional photometric characteristics, we propose to include polarization sensing. We demonstrate in this paper that the contribution of polarization measurement increases both the performances of segmentation and the quality of depth estimation. Our polarimetry-based approaches are compared here with other state-of-the-art RGB-centric methods showing interest of using polarization imaging.
Light emitted from a source into a scene can undergo complex interactions with scene surfaces of different material types before being reflected. During this transport, every surface reflection is encoded in the properties of the photons that reach the detector, including time, direction, intensity, wavelength and polarization. Conventional imaging systems capture intensity by integrating over all other dimensions of the light, hiding this rich scene information. Existing methods are capable of untangling these measurements into their spatial and temporal dimensions, fueling geometric scene understanding tasks. However, examining material properties jointly with geometric properties is an open challenge that could enable unprecedented capabilities beyond geometric scene understanding, allowing for material-dependent scene understanding and imaging through complex transport. In this work, we close this gap, and propose a computational light transport imaging method that captures the spatially- and temporally-resolved complete polarimetric response of a scene. Our method hinges on a 7D tensor theory of light transport. We discover low-rank structure in the polarimetric tensor dimension and propose a data-driven rotating ellipsometry method that learns to exploit redundancy of polarimetric structure. We instantiate our theory with two prototypes: spatio-polarimetric imaging and coaxial temporal-polarimetric imaging. This allows us, for the first time, to decompose scene light transport into temporal, spatial, and complete polarimetric dimensions that unveil scene properties hidden to conventional methods. We validate the applicability of our method on diverse tasks, including shape reconstruction with subsurface scattering, seeing through scattering media, untangling multi-bounce light transport, breaking metamerism, and decomposition of crystals.
Given a video captured from a first person perspective and the environment context of where the video is recorded, can we recognize what the person is doing and identify where the action occurs in the 3D space? We address this challenging problem of jointly recognizing and localizing actions of a mobile user on a known 3D map from egocentric videos. To this end, we propose a novel deep probabilistic model. Our model takes the inputs of a Hierarchical Volumetric Representation (HVR) of the 3D environment and an egocentric video, infers the 3D action location as a latent variable, and recognizes the action based on the video and contextual cues surrounding its potential locations. To evaluate our model, we conduct extensive experiments on the subset of Ego4D dataset, in which both human naturalistic actions and photo-realistic 3D environment reconstructions are captured. Our method demonstrates strong results on both action recognition and 3D action localization across seen and unseen environments. We believe our work points to an exciting research direction in the intersection of egocentric vision, and 3D scene understanding.
Extracting detailed 3D information of objects from video data is an important goal for holistic scene understanding. While recent methods have shown impressive results when reconstructing meshes of objects from a single image, results often remain ambiguous as part of the object is unobserved. Moreover, existing image-based datasets for mesh reconstruction don't permit to study models which integrate temporal information. To alleviate both concerns we present SAIL-VOS 3D: a synthetic video dataset with frame-by-frame mesh annotations which extends SAIL-VOS. We also develop first baselines for reconstruction of 3D meshes from video data via temporal models. We demonstrate efficacy of the proposed baseline on SAIL-VOS 3D and Pix3D, showing that temporal information improves reconstruction quality. Resources and additional information are available at http://sailvos.web.illinois.edu.
Scene understanding requires the extraction and representation of scene components together with their properties and inter-relations. We describe a model in which meaningful scene structures are extracted from the image by an iterative process, combining bottom-up (BU) and top-down (TD) networks, interacting through a symmetric bi-directional communication between them (counter-streams structure). The model constructs a scene representation by the iterative use of three components. The first model component is a BU stream that extracts selected scene elements, properties and relations. The second component (cognitive augmentation) augments the extracted visual representation based on relevant non-visual stored representations. It also provides input to the third component, the TD stream, in the form of a TD instruction, instructing the model what task to perform next. The TD stream then guides the BU visual stream to perform the selected task in the next cycle. During this process, the visual representations extracted from the image can be combined with relevant non-visual representations, so that the final scene representation is based on both visual information extracted from the scene and relevant stored knowledge of the world. We describe how a sequence of TD-instructions is used to extract from the scene structures of interest, including an algorithm to automatically select the next TD-instruction in the sequence. The extraction process is shown to have favorable properties in terms of combinatorial generalization, generalizing well to novel scene structures and new combinations of objects, properties and relations not seen during training. Finally, we compare the model with relevant aspects of the human vision, and suggest directions for using the BU-TD scheme for integrating visual and cognitive components in the process of scene understanding.
To detect and segment objects in images based on their content is one of the most active topics in the field of computer vision. Nowadays, this problem can be addressed using Deep Learning architectures such as Faster R-CNN or YOLO, among others. In this paper, we study the behaviour of different configurations of RetinaNet, Faster R-CNN and Mask R-CNN presented in Detectron2. First, we evaluate qualitatively and quantitatively (AP) the performance of the pre-trained models on KITTI-MOTS and MOTSChallenge datasets. We observe a significant improvement in performance after fine-tuning these models on the datasets of interest and optimizing hyperparameters. Finally, we run inference in unusual situations using out of context datasets, and present interesting results that help us understanding better the networks.
Lane-level scene annotations provide invaluable data in autonomous vehicles for trajectory planning in complex environments such as urban areas and cities. However, obtaining such data is time-consuming and expensive since lane annotations have to be annotated manually by humans and are as such hard to scale to large areas. In this work, we propose a novel approach for lane geometry estimation from bird's-eye-view images. We formulate the problem of lane shape and lane connections estimation as a graph estimation problem where lane anchor points are graph nodes and lane segments are graph edges. We train a graph estimation model on multimodal bird's-eye-view data processed from the popular NuScenes dataset and its map expansion pack. We furthermore estimate the direction of the lane connection for each lane segment with a separate model which results in a directed lane graph. We illustrate the performance of our LaneGraphNet model on the challenging NuScenes dataset and provide extensive qualitative and quantitative evaluation. Our model shows promising performance for most evaluated urban scenes and can serve as a step towards automated generation of HD lane annotations for autonomous driving.
Level-5 driving automation requires a robust visual perception system that can parse input images under any condition. However, existing driving datasets for dense semantic perception are either dominated by images captured under normal conditions or are small in scale. To address this, we introduce ACDC, the Adverse Conditions Dataset with Correspondences for training and testing methods for diverse semantic perception tasks on adverse visual conditions. ACDC consists of a large set of 8012 images, half of which (4006) are equally distributed between four common adverse conditions: fog, nighttime, rain, and snow. Each adverse-condition image comes with a high-quality pixel-level panoptic annotation, a corresponding image of the same scene under normal conditions, and a binary mask that distinguishes between intra-image regions of clear and uncertain semantic content. 1503 of the corresponding normal-condition images feature panoptic annotations, raising the total annotated images to 5509. ACDC supports the standard tasks of semantic segmentation, object detection, instance segmentation, and panoptic segmentation, as well as the newly introduced uncertainty-aware semantic segmentation. A detailed empirical study demonstrates the challenges that the adverse domains of ACDC pose to state-of-the-art supervised and unsupervised approaches and indicates the value of our dataset in steering future progress in the field. Our dataset and benchmark are publicly available at https://acdc.vision.ee.ethz.ch
The visual relationship recognition (VRR) task aims at understanding the pairwise visual relationships between interacting objects in an image. These relationships typically have a long-tail distribution due to their compositional nature. This problem gets more severe when the vocabulary becomes large, rendering this task very challenging. This paper shows that modeling an effective message-passing flow through an attention mechanism can be critical to tackling the compositionality and long-tail challenges in VRR. The method, called RelTransformer, represents each image as a fully-connected scene graph and restructures the whole scene into the relation-triplet and global-scene contexts. It directly passes the message from each element in the relation-triplet and global-scene contexts to the target relation via self-attention. We also design a learnable memory to augment the long-tail relation representation learning. Through extensive experiments, we find that our model generalizes well on many VRR benchmarks. Our model outperforms the best-performing models on two large-scale long-tail VRR benchmarks, VG8K-LT (+2.0% overall acc) and GQA-LT (+26.0% overall acc), both having a highly skewed distribution towards the tail. It also achieves strong results on the VG200 relation detection task. Our code is available at https://github.com/Vision-CAIR/RelTransformer.
Aerial scene recognition is a fundamental visual task and has attracted an increasing research interest in the last few years. Most of current researches mainly deploy efforts to categorize an aerial image into one scene-level label, while in real-world scenarios, there often exist multiple scenes in a single image. Therefore, in this paper, we propose to take a step forward to a more practical and challenging task, namely multi-scene recognition in single images. Moreover, we note that manually yielding annotations for such a task is extraordinarily time- and labor-consuming. To address this, we propose a prototype-based memory network to recognize multiple scenes in a single image by leveraging massive well-annotated single-scene images. The proposed network consists of three key components: 1) a prototype learning module, 2) a prototype-inhabiting external memory, and 3) a multi-head attention-based memory retrieval module. To be more specific, we first learn the prototype representation of each aerial scene from single-scene aerial image datasets and store it in an external memory. Afterwards, a multi-head attention-based memory retrieval module is devised to retrieve scene prototypes relevant to query multi-scene images for final predictions. Notably, only a limited number of annotated multi-scene images are needed in the training phase. To facilitate the progress of aerial scene recognition, we produce a new multi-scene aerial image (MAI) dataset. Experimental results on variant dataset configurations demonstrate the effectiveness of our network. Our dataset and codes are publicly available.
Human motion recognition (HMR) based on wireless sensing is a low-cost technique for scene understanding. Current HMR systems adopt support vector machines (SVMs) and convolutional neural networks (CNNs) to classify radar signals. However, whether a deeper learning model could improve the system performance is currently not known. On the other hand, training a machine learning model requires a large dataset, but data gathering from experiment is cost-expensive and time-consuming. Although wireless channel models can be adopted for dataset generation, current channel models are mostly designed for communication rather than sensing. To address the above problems, this paper proposes a deep spectrogram network (DSN) by leveraging the residual mapping technique to enhance the HMR performance. Furthermore, a primitive based autoregressive hybrid (PBAH) channel model is developed, which facilitates efficient training and testing dataset generation for HMR in a virtual environment. Experimental results demonstrate that the proposed PBAH channel model matches the actual experimental data very well and the proposed DSN achieves significantly smaller recognition error than that of CNN.
Detecting and localizing objects in the real 3D space, which plays a crucial role in scene understanding, is particularly challenging given only a monocular image due to the geometric information loss during imagery projection. We propose MonoGRNet for the amodal 3D object detection from a monocular image via geometric reasoning in both the observed 2D projection and the unobserved depth dimension. MonoGRNet decomposes the monocular 3D object detection task into four sub-tasks including 2D object detection, instance-level depth estimation, projected 3D center estimation and local corner regression. The task decomposition significantly facilitates the monocular 3D object detection, allowing the target 3D bounding boxes to be efficiently predicted in a single forward pass, without using object proposals, post-processing or the computationally expensive pixel-level depth estimation utilized by previous methods. In addition, MonoGRNet flexibly adapts to both fully and weakly supervised learning, which improves the feasibility of our framework in diverse settings. Experiments are conducted on KITTI, Cityscapes and MS COCO datasets. Results demonstrate the promising performance of our framework in various scenarios.
Point cloud semantic segmentation is a crucial task in 3D scene understanding. Existing methods mainly focus on employing a large number of annotated labels for supervised semantic segmentation. Nonetheless, manually labeling such large point clouds for the supervised segmentation task is time-consuming. In order to reduce the number of annotated labels, we propose a semi-supervised semantic point cloud segmentation network, named SSPC-Net, where we train the semantic segmentation network by inferring the labels of unlabeled points from the few annotated 3D points. In our method, we first partition the whole point cloud into superpoints and build superpoint graphs to mine the long-range dependencies in point clouds. Based on the constructed superpoint graph, we then develop a dynamic label propagation method to generate the pseudo labels for the unsupervised superpoints. Particularly, we adopt a superpoint dropout strategy to dynamically select the generated pseudo labels. In order to fully exploit the generated pseudo labels of the unsupervised superpoints, we furthermore propose a coupled attention mechanism for superpoint feature embedding. Finally, we employ the cross-entropy loss to train the semantic segmentation network with the labels of the supervised superpoints and the pseudo labels of the unsupervised superpoints. Experiments on various datasets demonstrate that our semi-supervised segmentation method can achieve better performance than the current semi-supervised segmentation method with fewer annotated 3D points. Our code is available at https://github.com/MMCheng/SSPC-Net.
We review solutions to the problem of depth estimation, arguably the most important subtask in scene understanding. We focus on the single image depth estimation problem. Due to its properties, the single image depth estimation problem is currently best tackled with machine learning methods, most successfully with convolutional neural networks. We provide an overview of the field by examining key works. We examine non-deep learning approaches that mostly predate deep learning and utilize hand-crafted features and assumptions, and more recent works that mostly use deep learning techniques. The single image depth estimation problem is tackled first in a supervised fashion with absolute or relative depth information acquired from human or sensor-labeled data, or in an unsupervised way using unlabelled stereo images or video datasets. We also study multitask approaches that combine the depth estimation problem with related tasks such as semantic segmentation and surface normal estimation. Finally, we discuss investigations into the mechanisms, principles, and failure cases of contemporary solutions.
Existing scene understanding systems mainly focus on recognizing the visible parts of a scene, ignoring the intact appearance of physical objects in the real-world. Concurrently, image completion has aimed to create plausible appearance for the invisible regions, but requires a manual mask as input. In this work, we propose a higher-level scene understanding system to tackle both visible and invisible parts of objects and backgrounds in a given scene. Particularly, we built a system to decompose a scene into individual objects, infer their underlying occlusion relationships, and even automatically learn which parts of the objects are occluded that need to be completed. In order to disentangle the occluded relationships of all objects in a complex scene, we use the fact that the front object without being occluded is easy to be identified, detected, and segmented. Our system interleaves the two tasks of instance segmentation and scene completion through multiple iterations, solving for objects layer-by-layer. We first provide a thorough experiment using a new realistically rendered dataset with ground-truths for all invisible regions. To bridge the domain gap to real imagery where ground-truths are unavailable, we then train another model with the pseudo-ground-truths generated from our trained synthesis model. We demonstrate results on a wide variety of datasets and show significant improvement over the state-of-the-art.
Semantic Scene Completion aims at reconstructing a complete 3D scene with precise voxel-wise semantics from a single-view depth or RGBD image. It is a crucial but challenging problem for indoor scene understanding. In this work, we present a novel framework named Scene-Instance-Scene Network (\textit{SISNet}), which takes advantages of both instance and scene level semantic information. Our method is capable of inferring fine-grained shape details as well as nearby objects whose semantic categories are easily mixed-up. The key insight is that we decouple the instances from a coarsely completed semantic scene instead of a raw input image to guide the reconstruction of instances and the overall scene. SISNet conducts iterative scene-to-instance (SI) and instance-to-scene (IS) semantic completion. Specifically, the SI is able to encode objects' surrounding context for effectively decoupling instances from the scene and each instance could be voxelized into higher resolution to capture finer details. With IS, fine-grained instance information can be integrated back into the 3D scene and thus leads to more accurate semantic scene completion. Utilizing such an iterative mechanism, the scene and instance completion benefits each other to achieve higher completion accuracy. Extensively experiments show that our proposed method consistently outperforms state-of-the-art methods on both real NYU, NYUCAD and synthetic SUNCG-RGBD datasets. The code and the supplementary material will be available at \url{https://github.com/yjcaimeow/SISNet}.