Scene understanding has made tremendous progress over the past few years, as data acquisition systems are now providing an increasing amount of data of various modalities (point cloud, depth, RGB...). However, this improvement comes at a large cost on computation resources and data annotation requirements. To analyze geometric information and images jointly, many approaches rely on both a 2D loss and 3D loss, requiring not only 2D per pixel-labels but also 3D per-point labels. However, obtaining a 3D groundtruth is challenging, time-consuming and error-prone. In this paper, we show that image segmentation can benefit from 3D geometric information without requiring a 3D groundtruth, by training the geometric feature extraction and the 2D segmentation network jointly, in an end-to-end fashion, using only the 2D segmentation loss. Our method starts by extracting a map of 3D features directly from a provided point cloud by using a lightweight 3D neural network. The 3D feature map, merged with the RGB image, is then used as an input to a classical image segmentation network. Our method can be applied to many 2D segmentation networks, improving significantly their performance with only a marginal network weight increase and light input dataset requirements, since no 3D groundtruth is required.
Most benchmarks for studying surgical interventions focus on a specific challenge instead of leveraging the intrinsic complementarity among different tasks. In this work, we present a new experimental framework towards holistic surgical scene understanding. First, we introduce the Phase, Step, Instrument, and Atomic Visual Action recognition (PSI-AVA) Dataset. PSI-AVA includes annotations for both long-term (Phase and Step recognition) and short-term reasoning (Instrument detection and novel Atomic Action recognition) in robot-assisted radical prostatectomy videos. Second, we present Transformers for Action, Phase, Instrument, and steps Recognition (TAPIR) as a strong baseline for surgical scene understanding. TAPIR leverages our dataset's multi-level annotations as it benefits from the learned representation on the instrument detection task to improve its classification capacity. Our experimental results in both PSI-AVA and other publicly available databases demonstrate the adequacy of our framework to spur future research on holistic surgical scene understanding.
Image instance segmentation is a fundamental research topic in autonomous driving, which is crucial for scene understanding and road safety. Advanced learning-based approaches often rely on the costly 2D mask annotations for training. In this paper, we present a more artful framework, LiDAR-guided Weakly Supervised Instance Segmentation (LWSIS), which leverages the off-the-shelf 3D data, i.e., Point Cloud, together with the 3D boxes, as natural weak supervisions for training the 2D image instance segmentation models. Our LWSIS not only exploits the complementary information in multimodal data during training, but also significantly reduces the annotation cost of the dense 2D masks. In detail, LWSIS consists of two crucial modules, Point Label Assignment (PLA) and Graph-based Consistency Regularization (GCR). The former module aims to automatically assign the 3D point cloud as 2D point-wise labels, while the latter further refines the predictions by enforcing geometry and appearance consistency of the multimodal data. Moreover, we conduct a secondary instance segmentation annotation on the nuScenes, named nuInsSeg, to encourage further research on multimodal perception tasks. Extensive experiments on the nuInsSeg, as well as the large-scale Waymo, show that LWSIS can substantially improve existing weakly supervised segmentation models by only involving 3D data during training. Additionally, LWSIS can also be incorporated into 3D object detectors like PointPainting to boost the 3D detection performance for free. The code and dataset are available at https://github.com/Serenos/LWSIS.
Scene understanding is an essential and challenging task in computer vision. To provide the visually fundamental graphical structure of an image, the scene graph has received increased attention due to its powerful semantic representation. However, it is difficult to draw a proper scene graph for image retrieval, image generation, and multi-modal applications. The conventional scene graph annotation interface is not easy to use in image annotations, and the automatic scene graph generation approaches using deep neural networks are prone to generate redundant content while disregarding details. In this work, we propose SGDraw, a scene graph drawing interface using object-oriented scene graph representation to help users draw and edit scene graphs interactively. For the proposed object-oriented representation, we consider the objects, attributes, and relationships of objects as a structural unit. SGDraw provides a web-based scene graph annotation and generation tool for scene understanding applications. To verify the effectiveness of the proposed interface, we conducted a comparison study with the conventional tool and the user experience study. The results show that SGDraw can help generate scene graphs with richer details and describe the images more accurately than traditional bounding box annotations. We believe the proposed SGDraw can be useful in various vision tasks, such as image retrieval and generation.
Open-vocabulary scene understanding aims to localize and recognize unseen categories beyond the annotated label space. The recent breakthrough of 2D open-vocabulary perception is largely driven by Internet-scale paired image-text data with rich vocabulary concepts. However, this success cannot be directly transferred to 3D scenarios due to the inaccessibility of large-scale 3D-text pairs. To this end, we propose to distill knowledge encoded in pre-trained vision-language (VL) foundation models through captioning multi-view images from 3D, which allows explicitly associating 3D and semantic-rich captions. Further, to foster coarse-to-fine visual-semantic representation learning from captions, we design hierarchical 3D-caption pairs, leveraging geometric constraints between 3D scenes and multi-view images. Finally, by employing contrastive learning, the model learns language-aware embeddings that connect 3D and text for open-vocabulary tasks. Our method not only remarkably outperforms baseline methods by 25.8% $\sim$ 44.7% hIoU and 14.5% $\sim$ 50.4% hAP$_{50}$ in open-vocabulary semantic and instance segmentation, but also shows robust transferability on challenging zero-shot domain transfer tasks. See the project website at https://dingry.github.io/projects/PLA.
Purpose: Surgery scene understanding with tool-tissue interaction recognition and automatic report generation can play an important role in intra-operative guidance, decision-making and postoperative analysis in robotic surgery. However, domain shifts between different surgeries with inter and intra-patient variation and novel instruments' appearance degrade the performance of model prediction. Moreover, it requires output from multiple models, which can be computationally expensive and affect real-time performance. Methodology: A multi-task learning (MTL) model is proposed for surgical report generation and tool-tissue interaction prediction that deals with domain shift problems. The model forms of shared feature extractor, mesh-transformer branch for captioning and graph attention branch for tool-tissue interaction prediction. The shared feature extractor employs class incremental contrastive learning (CICL) to tackle intensity shift and novel class appearance in the target domain. We design Laplacian of Gaussian (LoG) based curriculum learning into both shared and task-specific branches to enhance model learning. We incorporate a task-aware asynchronous MTL optimization technique to fine-tune the shared weights and converge both tasks optimally. Results: The proposed MTL model trained using task-aware optimization and fine-tuning techniques reported a balanced performance (BLEU score of 0.4049 for scene captioning and accuracy of 0.3508 for interaction detection) for both tasks on the target domain and performed on-par with single-task models in domain adaptation. Conclusion: The proposed multi-task model was able to adapt to domain shifts, incorporate novel instruments in the target domain, and perform tool-tissue interaction detection and report generation on par with single-task models.
Learning descriptive 3D features is crucial for understanding 3D scenes with diverse objects and complex structures. However, it is usually unknown whether important geometric attributes and scene context obtain enough emphasis in an end-to-end trained 3D scene understanding network. To guide 3D feature learning toward important geometric attributes and scene context, we explore the help of textual scene descriptions. Given some free-form descriptions paired with 3D scenes, we extract the knowledge regarding the object relationships and object attributes. We then inject the knowledge to 3D feature learning through three classification-based auxiliary tasks. This language-assisted training can be combined with modern object detection and instance segmentation methods to promote 3D semantic scene understanding, especially in a label-deficient regime. Moreover, the 3D feature learned with language assistance is better aligned with the language features, which can benefit various 3D-language multimodal tasks. Experiments on several benchmarks of 3D-only and 3D-language tasks demonstrate the effectiveness of our language-assisted 3D feature learning. Code is available at https://github.com/Asterisci/Language-Assisted-3D.
Road scene understanding tasks have recently become crucial for self-driving vehicles. In particular, real-time semantic segmentation is indispensable for intelligent self-driving agents to recognize roadside objects in the driving area. As prior research works have primarily sought to improve the segmentation performance with computationally heavy operations, they require far significant hardware resources for both training and deployment, and thus are not suitable for real-time applications. As such, we propose a doubly contrastive approach to improve the performance of a more practical lightweight model for self-driving, specifically under adverse weather conditions such as fog, nighttime, rain and snow. Our proposed approach exploits both image- and pixel-level contrasts in an end-to-end supervised learning scheme without requiring a memory bank for global consistency or the pretraining step used in conventional contrastive methods. We validate the effectiveness of our method using SwiftNet on the ACDC dataset, where it achieves up to 1.34%p improvement in mIoU (ResNet-18 backbone) at 66.7 FPS (2048x1024 resolution) on a single RTX 3080 Mobile GPU at inference. Furthermore, we demonstrate that replacing image-level supervision with self-supervision achieves comparable performance when pre-trained with clear weather images.
Semantic scene understanding with Minimalist Optical Systems (MOS) in mobile and wearable applications remains a challenge due to the corrupted imaging quality induced by optical aberrations. However, previous works only focus on improving the subjective imaging quality through the Computational Imaging (CI) technique, ignoring the feasibility of advancing semantic segmentation. In this paper, we pioneer the investigation of Semantic Segmentation under Optical Aberrations (SSOA) with MOS. To benchmark SSOA, we construct Virtual Prototype Lens (VPL) groups through optical simulation, generating Cityscapes-ab and KITTI-360-ab datasets under different behaviors and levels of aberrations. We look into SSOA via an unsupervised domain adaptation perspective to address the scarcity of labeled aberration data in real-world scenarios. Further, we propose Computational Imaging Assisted Domain Adaptation (CIADA) to leverage prior knowledge of CI for robust performance in SSOA. Based on our benchmark, we conduct experiments on the robustness of classical segmenters against aberrations. In addition, extensive evaluations of possible solutions to SSOA reveal that CIADA achieves superior performance under all aberration distributions, bridging the gap between computational imaging and downstream applications for MOS. The project page is at https://github.com/zju-jiangqi/CIADA.
3D object detection from multiple image views is a fundamental and challenging task for visual scene understanding. Owing to its low cost and high efficiency, multi-view 3D object detection has demonstrated promising application prospects. However, accurately detecting objects through perspective views is extremely difficult due to the lack of depth information. Current approaches tend to adopt heavy backbones for image encoders, making them inapplicable for real-world deployment. Different from the images, LiDAR points are superior in providing spatial cues, resulting in highly precise localization. In this paper, we explore the incorporation of LiDAR-based detectors for multi-view 3D object detection. Instead of directly training a depth prediction network, we unify the image and LiDAR features in the Bird-Eye-View (BEV) space and adaptively transfer knowledge across non-homogenous representations in a teacher-student paradigm. To this end, we propose \textbf{BEVDistill}, a cross-modal BEV knowledge distillation (KD) framework for multi-view 3D object detection. Extensive experiments demonstrate that the proposed method outperforms current KD approaches on a highly-competitive baseline, BEVFormer, without introducing any extra cost in the inference phase. Notably, our best model achieves 59.4 NDS on the nuScenes test leaderboard, achieving new state-of-the-art in comparison with various image-based detectors. Code will be available at https://github.com/zehuichen123/BEVDistill.
Most recent work in visual sound source localization relies on semantic audio-visual representations learned in a self-supervised manner, and by design excludes temporal information present in videos. While it proves to be effective for widely used benchmark datasets, the method falls short for challenging scenarios like urban traffic. This work introduces temporal context into the state-of-the-art methods for sound source localization in urban scenes using optical flow as a means to encode motion information. An analysis of the strengths and weaknesses of our methods helps us better understand the problem of visual sound source localization and sheds light on open challenges for audio-visual scene understanding.
Visual Grounding (VG) in Visual Question Answering (VQA) systems describes how well a system manages to tie a question and its answer to relevant image regions. Systems with strong VG are considered intuitively interpretable and suggest an improved scene understanding. While VQA accuracy performances have seen impressive gains over the past few years, explicit improvements to VG performance and evaluation thereof have often taken a back seat on the road to overall accuracy improvements. A cause of this originates in the predominant choice of learning paradigm for VQA systems, which consists of training a discriminative classifier over a predetermined set of answer options. In this work, we break with the dominant VQA modeling paradigm of classification and investigate VQA from the standpoint of an information retrieval task. As such, the developed system directly ties VG into its core search procedure. Our system operates over a weighted, directed, acyclic graph, a.k.a. "lattice", which is derived from the scene graph of a given image in conjunction with region-referring expressions extracted from the question. We give a detailed analysis of our approach and discuss its distinctive properties and limitations. Our approach achieves the strongest VG performance among examined systems and exhibits exceptional generalization capabilities in a number of scenarios.
Autonomous driving is complex, requiring sophisticated 3D scene understanding, localization, mapping, and control. Rather than explicitly modelling and fusing each of these components, we instead consider an end-to-end approach via reinforcement learning (RL). However, collecting exploration driving data in the real world is impractical and dangerous. While training in simulation and deploying visual sim-to-real techniques has worked well for robot manipulation, deploying beyond controlled workspace viewpoints remains a challenge. In this paper, we address this challenge by presenting Sim2Seg, a re-imagining of RCAN that crosses the visual reality gap for off-road autonomous driving, without using any real-world data. This is done by learning to translate randomized simulation images into simulated segmentation and depth maps, subsequently enabling real-world images to also be translated. This allows us to train an end-to-end RL policy in simulation, and directly deploy in the real-world. Our approach, which can be trained in 48 hours on 1 GPU, can perform equally as well as a classical perception and control stack that took thousands of engineering hours over several months to build. We hope this work motivates future end-to-end autonomous driving research.
Semantic segmentation is important for scene understanding. To address the scenes of adverse illumination conditions of natural images, thermal infrared (TIR) images are introduced. Most existing RGB-T semantic segmentation methods follow three cross-modal fusion paradigms, i.e. encoder fusion, decoder fusion, and feature fusion. Some methods, unfortunately, ignore the properties of RGB and TIR features or the properties of features at different levels. In this paper, we propose a novel feature fusion-based network for RGB-T semantic segmentation, named \emph{LASNet}, which follows three steps of location, activation, and sharpening. The highlight of LASNet is that we fully consider the characteristics of cross-modal features at different levels, and accordingly propose three specific modules for better segmentation. Concretely, we propose a Collaborative Location Module (CLM) for high-level semantic features, aiming to locate all potential objects. We propose a Complementary Activation Module for middle-level features, aiming to activate exact regions of different objects. We propose an Edge Sharpening Module (ESM) for low-level texture features, aiming to sharpen the edges of objects. Furthermore, in the training phase, we attach a location supervision and an edge supervision after CLM and ESM, respectively, and impose two semantic supervisions in the decoder part to facilitate network convergence. Experimental results on two public datasets demonstrate that the superiority of our LASNet over relevant state-of-the-art methods. The code and results of our method are available at https://github.com/MathLee/LASNet.
In Multi-Task Learning (MTL), tasks may compete and limit the performance achieved on each other, rather than guiding the optimization to a solution, superior to all its single-task trained counterparts. Since there is often not a unique solution optimal for all tasks, practitioners have to balance tradeoffs between tasks' performance, and resort to optimality in the Pareto sense. Most MTL methodologies either completely neglect this aspect, and instead of aiming at learning a Pareto Front, produce one solution predefined by their optimization schemes, or produce diverse but discrete solutions. Recent approaches parameterize the Pareto Front via neural networks, leading to complex mappings from tradeoff to objective space. In this paper, we conjecture that the Pareto Front admits a linear parameterization in parameter space, which leads us to propose \textit{Pareto Manifold Learning}, an ensembling method in weight space. Our approach produces a continuous Pareto Front in a single training run, that allows to modulate the performance on each task during inference. Experiments on multi-task learning benchmarks, ranging from image classification to tabular datasets and scene understanding, show that \textit{Pareto Manifold Learning} outperforms state-of-the-art single-point algorithms, while learning a better Pareto parameterization than multi-point baselines.
We propose a new task to benchmark scene understanding of embodied agents: Situated Question Answering in 3D Scenes (SQA3D). Given a scene context (e.g., 3D scan), SQA3D requires the tested agent to first understand its situation (position, orientation, etc.) in the 3D scene as described by text, then reason about its surrounding environment and answer a question under that situation. Based upon 650 scenes from ScanNet, we provide a dataset centered around 6.8k unique situations, along with 20.4k descriptions and 33.4k diverse reasoning questions for these situations. These questions examine a wide spectrum of reasoning capabilities for an intelligent agent, ranging from spatial relation comprehension to commonsense understanding, navigation, and multi-hop reasoning. SQA3D imposes a significant challenge to current multi-modal especially 3D reasoning models. We evaluate various state-of-the-art approaches and find that the best one only achieves an overall score of 47.20%, while amateur human participants can reach 90.06%. We believe SQA3D could facilitate future embodied AI research with stronger situation understanding and reasoning capability.
Generating a 3D point cloud from a single 2D image is of great importance for 3D scene understanding applications. To reconstruct the whole 3D shape of the object shown in the image, the existing deep learning based approaches use either explicit or implicit generative modeling of point clouds, which, however, suffer from limited quality. In this work, we aim to alleviate this issue by introducing a hybrid explicit-implicit generative modeling scheme, which inherits the flow-based explicit generative models for sampling point clouds with arbitrary resolutions while improving the detailed 3D structures of point clouds by leveraging the implicit generative adversarial networks (GANs). We evaluate on the large-scale synthetic dataset ShapeNet, with the experimental results demonstrating the superior performance of the proposed method. In addition, the generalization ability of our method is demonstrated by performing on cross-category synthetic images as well as by testing on real images from PASCAL3D+ dataset.
Modern autonomous systems often rely on LiDAR scanners, in particular for autonomous driving scenarios. In this context, reliable scene understanding is indispensable. Current learning-based methods typically try to achieve maximum performance for this task, while neglecting a proper estimation of the associated uncertainties. In this work, we introduce a novel approach for solving the task of uncertainty-aware panoptic segmentation using LiDAR point clouds. Our proposed EvLPSNet network is the first to solve this task efficiently in a sampling-free manner. It aims to predict per-point semantic and instance segmentations, together with per-point uncertainty estimates. Moreover, it incorporates methods for improving the performance by employing the predicted uncertainties. We provide several strong baselines combining state-of-the-art panoptic segmentation networks with sampling-free uncertainty estimation techniques. Extensive evaluations show that we achieve the best performance on uncertainty-aware panoptic segmentation quality and calibration compared to these baselines. We make our code available at: \url{https://github.com/kshitij3112/EvLPSNet}
Self-supervised monocular depth estimation is a salient task for 3D scene understanding. Learned jointly with monocular ego-motion estimation, several methods have been proposed to predict accurate pixel-wise depth without using labeled data. Nevertheless, these methods focus on improving performance under ideal conditions without natural or digital corruptions. The general absence of occlusions is assumed even for object-specific depth estimation. These methods are also vulnerable to adversarial attacks, which is a pertinent concern for their reliable deployment in robots and autonomous driving systems. We propose MIMDepth, a method that adapts masked image modeling (MIM) for self-supervised monocular depth estimation. While MIM has been used to learn generalizable features during pre-training, we show how it could be adapted for direct training of monocular depth estimation. Our experiments show that MIMDepth is more robust to noise, blur, weather conditions, digital artifacts, occlusions, as well as untargeted and targeted adversarial attacks.
In this work we present FreDSNet, a deep learning solution which obtains semantic 3D understanding of indoor environments from single panoramas. Omnidirectional images reveal task-specific advantages when addressing scene understanding problems due to the 360-degree contextual information about the entire environment they provide. However, the inherent characteristics of the omnidirectional images add additional problems to obtain an accurate detection and segmentation of objects or a good depth estimation. To overcome these problems, we exploit convolutions in the frequential domain obtaining a wider receptive field in each convolutional layer. These convolutions allow to leverage the whole context information from omnidirectional images. FreDSNet is the first network that jointly provides monocular depth estimation and semantic segmentation from a single panoramic image exploiting fast Fourier convolutions. Our experiments show that FreDSNet has similar performance as specific state of the art methods for semantic segmentation and depth estimation. FreDSNet code is publicly available in https://github.com/Sbrunoberenguel/FreDSNet