Federico Tombari

Tao Sun, Mattia Segu, Janis Postels et al. · eth-zurich, stanford

Adapting to a continuously evolving environment is a safety-critical challenge inevitably faced by all autonomous driving systems. Existing image and video driving datasets, however, fall short of capturing the mutable nature of the real world. In this paper, we introduce the largest multi-task synthetic dataset for autonomous driving, SHIFT. It presents discrete and continuous shifts in cloudiness, rain and fog intensity, time of day, and vehicle and pedestrian density. Featuring a comprehensive sensor suite and annotations for several mainstream perception tasks, SHIFT allows investigating the degradation of a perception system performance at increasing levels of domain shift, fostering the development of continuous adaptation strategies to mitigate this problem and assess model robustness and generality. Our dataset and benchmark toolkit are publicly available at www.vis.xyz/shift.

Gu Wang, Fabian Manhardt, Xingyu Liu et al. · tsinghua

6D object pose estimation is a fundamental yet challenging problem in computer vision. Convolutional Neural Networks (CNNs) have recently proven to be capable of predicting reliable 6D pose estimates even under monocular settings. Nonetheless, CNNs are identified as being extremely data-driven, and acquiring adequate annotations is oftentimes very time-consuming and labor intensive. To overcome this limitation, we propose a novel monocular 6D pose estimation approach by means of self-supervised learning, removing the need for real annotations. After training our proposed network fully supervised with synthetic RGB data, we leverage current trends in noisy student training and differentiable rendering to further self-supervise the model on these unsupervised real RGB(-D) samples, seeking for a visually and geometrically optimal alignment. Moreover, employing both visible and amodal mask information, our self-supervision becomes very robust towards challenging scenarios such as occlusion. Extensive evaluations demonstrate that our proposed self-supervision outperforms all other methods relying on synthetic data or employing elaborate techniques from the domain adaptation realm. Noteworthy, our self-supervised approach consistently improves over its synthetically trained baseline and often almost closes the gap towards its fully supervised counterpart. The code and models are publicly available at https://github.com/THU-DA-6D-Pose-Group/self6dpp.git.

Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona et al.

By identifying four important components of existing LiDAR-camera 3D object detection methods (LiDAR and camera candidates, transformation, and fusion outputs), we observe that all existing methods either find dense candidates or yield dense representations of scenes. However, given that objects occupy only a small part of a scene, finding dense candidates and generating dense representations is noisy and inefficient. We propose SparseFusion, a novel multi-sensor 3D detection method that exclusively uses sparse candidates and sparse representations. Specifically, SparseFusion utilizes the outputs of parallel detectors in the LiDAR and camera modalities as sparse candidates for fusion. We transform the camera candidates into the LiDAR coordinate space by disentangling the object representations. Then, we can fuse the multi-modality candidates in a unified 3D space by a lightweight self-attention module. To mitigate negative transfer between modalities, we propose novel semantic and geometric cross-modality transfer modules that are applied prior to the modality-specific detectors. SparseFusion achieves state-of-the-art performance on the nuScenes benchmark while also running at the fastest speed, even outperforming methods with stronger backbones. We perform extensive experiments to demonstrate the effectiveness and efficiency of our modules and overall method pipeline. Our code will be made publicly available at https://github.com/yichen928/SparseFusion.

Guangyao Zhai, Dianye Huang, Shun-Cheng Wu et al.

6-DoF robotic grasping is a long-lasting but unsolved problem. Recent methods utilize strong 3D networks to extract geometric grasping representations from depth sensors, demonstrating superior accuracy on common objects but perform unsatisfactorily on photometrically challenging objects, e.g., objects in transparent or reflective materials. The bottleneck lies in that the surface of these objects can not reflect back accurate depth due to the absorption or refraction of light. In this paper, in contrast to exploiting the inaccurate depth data, we propose the first RGB-only 6-DoF grasping pipeline called MonoGraspNet that utilizes stable 2D features to simultaneously handle arbitrary object grasping and overcome the problems induced by photometrically challenging objects. MonoGraspNet leverages keypoint heatmap and normal map to recover the 6-DoF grasping poses represented by our novel representation parameterized with 2D keypoints with corresponding depth, grasping direction, grasping width, and angle. Extensive experiments in real scenes demonstrate that our method can achieve competitive results in grasping common objects and surpass the depth-based competitor by a large margin in grasping photometrically challenging objects. To further stimulate robotic manipulation research, we additionally annotate and open-source a multi-view and multi-scene real-world grasping dataset, containing 120 objects of mixed photometric complexity with 20M accurate grasping labels.

Lukas Schmid, Mansoor Nasir Cheema, Victor Reijgwart et al.

Exploration of unknown environments is a fundamental problem in robotics and an essential component in numerous applications of autonomous systems. A major challenge in exploring unknown environments is that the robot has to plan with the limited information available at each time step. While most current approaches rely on heuristics and assumption to plan paths based on these partial observations, we instead propose a novel way to integrate deep learning into exploration by leveraging 3D scene completion for informed, safe, and interpretable exploration mapping and planning. Our approach, SC-Explorer, combines scene completion using a novel incremental fusion mechanism and a newly proposed hierarchical multi-layer mapping approach, to guarantee safety and efficiency of the robot. We further present an informative path planning method, leveraging the capabilities of our mapping approach and a novel scene-completion-aware information gain. While our method is generally applicable, we evaluate it in the use case of a Micro Aerial Vehicle (MAV). We thoroughly study each component in high-fidelity simulation experiments using only mobile hardware, and show that our method can speed up coverage of an environment by 73% compared to the baselines with only minimal reduction in map accuracy. Even if scene completions are not included in the final map, we show that they can be used to guide the robot to choose more informative paths, speeding up the measurement of the scene with the robot's sensors by 35%. We validate our system on a fully autonomous MAV, showing rapid and reliable scene coverage even in a complex cluttered environment. We make our methods available as open-source.

Evin Pınar Örnek, Aravindhan K Krishnan, Shreekant Gayaka et al.

Object instance segmentation is a key challenge for indoor robots navigating cluttered environments with many small objects. Limitations in 3D sensing capabilities often make it difficult to detect every possible object. While deep learning approaches may be effective for this problem, manually annotating 3D data for supervised learning is time-consuming. In this work, we explore zero-shot instance segmentation (ZSIS) from RGB-D data to identify unseen objects in a semantic category-agnostic manner. We introduce a zero-shot split for Tabletop Objects Dataset (TOD-Z) to enable this study and present a method that uses annotated objects to learn the ``objectness'' of pixels and generalize to unseen object categories in cluttered indoor environments. Our method, SupeRGB-D, groups pixels into small patches based on geometric cues and learns to merge the patches in a deep agglomerative clustering fashion. SupeRGB-D outperforms existing baselines on unseen objects while achieving similar performance on seen objects. We further show competitive results on the real dataset OCID. With its lightweight design (0.4 MB memory requirement), our method is extremely suitable for mobile and robotic applications. Additional DINO features can increase performance with a higher memory requirement. The dataset split and code are available at https://github.com/evinpinar/supergb-d.

Ayça Takmaz, Elisabetta Fedele, Robert W. Sumner et al.

We introduce the task of open-vocabulary 3D instance segmentation. Current approaches for 3D instance segmentation can typically only recognize object categories from a pre-defined closed set of classes that are annotated in the training datasets. This results in important limitations for real-world applications where one might need to perform tasks guided by novel, open-vocabulary queries related to a wide variety of objects. Recently, open-vocabulary 3D scene understanding methods have emerged to address this problem by learning queryable features for each point in the scene. While such a representation can be directly employed to perform semantic segmentation, existing methods cannot separate multiple object instances. In this work, we address this limitation, and propose OpenMask3D, which is a zero-shot approach for open-vocabulary 3D instance segmentation. Guided by predicted class-agnostic 3D instance masks, our model aggregates per-mask features via multi-view fusion of CLIP-based image embeddings. Experiments and ablation studies on ScanNet200 and Replica show that OpenMask3D outperforms other open-vocabulary methods, especially on the long-tail distribution. Qualitative experiments further showcase OpenMask3D's ability to segment object properties based on free-form queries describing geometry, affordances, and materials.

Chenyangguang Zhang, Yan Di, Ruida Zhang et al.

Reconstructing hand-held objects from a single RGB image is an important and challenging problem. Existing works utilizing Signed Distance Fields (SDF) reveal limitations in comprehensively capturing the complex hand-object interactions, since SDF is only reliable within the proximity of the target, and hence, infeasible to simultaneously encode local hand and object cues. To address this issue, we propose DDF-HO, a novel approach leveraging Directed Distance Field (DDF) as the shape representation. Unlike SDF, DDF maps a ray in 3D space, consisting of an origin and a direction, to corresponding DDF values, including a binary visibility signal determining whether the ray intersects the objects and a distance value measuring the distance from origin to target in the given direction. We randomly sample multiple rays and collect local to global geometric features for them by introducing a novel 2D ray-based feature aggregation scheme and a 3D intersection-aware hand pose embedding, combining 2D-3D features to model hand-object interactions. Extensive experiments on synthetic and real-world datasets demonstrate that DDF-HO consistently outperforms all baseline methods by a large margin, especially under Chamfer Distance, with about 80% leap forward. Codes are available at https://github.com/ZhangCYG/DDFHO.

Lukas Hoyer, David Joseph Tan, Muhammad Ferjad Naeem et al.

In semi-supervised semantic segmentation, a model is trained with a limited number of labeled images along with a large corpus of unlabeled images to reduce the high annotation effort. While previous methods are able to learn good segmentation boundaries, they are prone to confuse classes with similar visual appearance due to the limited supervision. On the other hand, vision-language models (VLMs) are able to learn diverse semantic knowledge from image-caption datasets but produce noisy segmentation due to the image-level training. In SemiVL, we propose to integrate rich priors from VLM pre-training into semi-supervised semantic segmentation to learn better semantic decision boundaries. To adapt the VLM from global to local reasoning, we introduce a spatial fine-tuning strategy for label-efficient learning. Further, we design a language-guided decoder to jointly reason over vision and language. Finally, we propose to handle inherent ambiguities in class labels by providing the model with language guidance in the form of class definitions. We evaluate SemiVL on 4 semantic segmentation datasets, where it significantly outperforms previous semi-supervised methods. For instance, SemiVL improves the state-of-the-art by +13.5 mIoU on COCO with 232 annotated images and by +6.1 mIoU on Pascal VOC with 92 labels. Project page: https://github.com/google-research/semivl

Andy Zeng, Maria Attarian, Brian Ichter et al.

Large pretrained (e.g., "foundation") models exhibit distinct capabilities depending on the domain of data they are trained on. While these domains are generic, they may only barely overlap. For example, visual-language models (VLMs) are trained on Internet-scale image captions, but large language models (LMs) are further trained on Internet-scale text with no images (e.g., spreadsheets, SAT questions, code). As a result, these models store different forms of commonsense knowledge across different domains. In this work, we show that this diversity is symbiotic, and can be leveraged through Socratic Models (SMs): a modular framework in which multiple pretrained models may be composed zero-shot i.e., via multimodal-informed prompting, to exchange information with each other and capture new multimodal capabilities, without requiring finetuning. With minimal engineering, SMs are not only competitive with state-of-the-art zero-shot image captioning and video-to-text retrieval, but also enable new applications such as (i) answering free-form questions about egocentric video, (ii) engaging in multimodal assistive dialogue with people (e.g., for cooking recipes) by interfacing with external APIs and databases (e.g., web search), and (iii) robot perception and planning.

Mahdi Saleh, Yige Wang, Nassir Navab et al.

Processing 3D data efficiently has always been a challenge. Spatial operations on large-scale point clouds, stored as sparse data, require extra cost. Attracted by the success of transformers, researchers are using multi-head attention for vision tasks. However, attention calculations in transformers come with quadratic complexity in the number of inputs and miss spatial intuition on sets like point clouds. We redesign set transformers in this work and incorporate them into a hierarchical framework for shape classification and part and scene segmentation. We propose our local attention unit, which captures features in a spatial neighborhood. We also compute efficient and dynamic global cross attentions by leveraging sampling and grouping at each iteration. Finally, to mitigate the non-heterogeneity of point clouds, we propose an efficient Multi-Scale Tokenization (MST), which extracts scale-invariant tokens for attention operations. The proposed hierarchical model achieves state-of-the-art shape classification in mean accuracy and yields results on par with the previous segmentation methods while requiring significantly fewer computations. Our proposed architecture predicts segmentation labels with around half the latency and parameter count of the previous most efficient method with comparable performance. The code is available at https://github.com/YigeWang-WHU/CloudAttention.

Muhammad Ferjad Naeem, Muhammad Gul Zain Ali Khan, Yongqin Xian et al.

Recent works have shown that unstructured text (documents) from online sources can serve as useful auxiliary information for zero-shot image classification. However, these methods require access to a high-quality source like Wikipedia and are limited to a single source of information. Large Language Models (LLM) trained on web-scale text show impressive abilities to repurpose their learned knowledge for a multitude of tasks. In this work, we provide a novel perspective on using an LLM to provide text supervision for a zero-shot image classification model. The LLM is provided with a few text descriptions from different annotators as examples. The LLM is conditioned on these examples to generate multiple text descriptions for each class(referred to as views). Our proposed model, I2MVFormer, learns multi-view semantic embeddings for zero-shot image classification with these class views. We show that each text view of a class provides complementary information allowing a model to learn a highly discriminative class embedding. Moreover, we show that I2MVFormer is better at consuming the multi-view text supervision from LLM compared to baseline models. I2MVFormer establishes a new state-of-the-art on three public benchmark datasets for zero-shot image classification with unsupervised semantic embeddings.

Muhammad Gul Zain Ali Khan, Muhammad Ferjad Naeem, Luc Van Gool et al.

Continual Learning aims to learn a single model on a sequence of tasks without having access to data from previous tasks. The biggest challenge in the domain still remains catastrophic forgetting: a loss in performance on seen classes of earlier tasks. Some existing methods rely on an expensive replay buffer to store a chunk of data from previous tasks. This, while promising, becomes expensive when the number of tasks becomes large or data can not be stored for privacy reasons. As an alternative, prompt-based methods have been proposed that store the task information in a learnable prompt pool. This prompt pool instructs a frozen image encoder on how to solve each task. While the model faces a disjoint set of classes in each task in this setting, we argue that these classes can be encoded to the same embedding space of a pre-trained language encoder. In this work, we propose Language Guidance for Prompt-based Continual Learning (LGCL) as a plug-in for prompt-based methods. LGCL is model agnostic and introduces language guidance at the task level in the prompt pool and at the class level on the output feature of the vision encoder. We show with extensive experimentation that LGCL consistently improves the performance of prompt-based continual learning methods to set a new state-of-the art. LGCL achieves these performance improvements without needing any additional learnable parameters.

Changxuan Li, Nadine Berner, Nassir Navab et al.

Self-supervised depth estimation from monocular sequences relies on the joint learning of a depth and a pose network. Despite abundant research done to improve the depth network, efforts on the pose remain limited. In this context, even when depth is estimated up to scale, we highlight the importance of the alignment between the scene scales estimated by the pose and depth nets. Then, we introduce SA4Depth, an approach to improve this alignment and boost the depth predictions while keeping the inference time unchanged. Our proposed method uses the depth estimated during training to reproject learnable visual features across consecutive frames and refine the pose estimates by reducing feature alignment residuals. With our method, the estimated scene scales by the separate depth and pose networks are aligned, and the prediction scale consistency is improved across different sequences. Our differentiable refinement integrates seamlessly into existing self-supervised pipelines and substantially improves their depth estimates. We demonstrate this with extensive experiments both outdoors and indoors on KITTI, Cityscapes, and NYUv2. Additionally, results on KITTI Odometry confirm the effectiveness of our pose refinement. Our code is available at https://github.com/Runningchauncey/SA4Depth .

Muhammad Ferjad Naeem, Yongqin Xian, Luc Van Gool et al.

Despite the tremendous progress in zero-shot learning(ZSL), the majority of existing methods still rely on human-annotated attributes, which are difficult to annotate and scale. An unsupervised alternative is to represent each class using the word embedding associated with its semantic class name. However, word embeddings extracted from pre-trained language models do not necessarily capture visual similarities, resulting in poor zero-shot performance. In this work, we argue that online textual documents, e.g., Wikipedia, contain rich visual descriptions about object classes, therefore can be used as powerful unsupervised side information for ZSL. To this end, we propose I2DFormer, a novel transformer-based ZSL framework that jointly learns to encode images and documents by aligning both modalities in a shared embedding space. In order to distill discriminative visual words from noisy documents, we introduce a new cross-modal attention module that learns fine-grained interactions between image patches and document words. Consequently, our I2DFormer not only learns highly discriminative document embeddings that capture visual similarities but also gains the ability to localize visually relevant words in image regions. Quantitatively, we demonstrate that our I2DFormer significantly outperforms previous unsupervised semantic embeddings under both zero-shot and generalized zero-shot learning settings on three public datasets. Qualitatively, we show that our method leads to highly interpretable results where document words can be grounded in the image regions.

Yongzhi Su, Mahdi Saleh, Torben Fetzer et al.

Establishing correspondences from image to 3D has been a key task of 6DoF object pose estimation for a long time. To predict pose more accurately, deeply learned dense maps replaced sparse templates. Dense methods also improved pose estimation in the presence of occlusion. More recently researchers have shown improvements by learning object fragments as segmentation. In this work, we present a discrete descriptor, which can represent the object surface densely. By incorporating a hierarchical binary grouping, we can encode the object surface very efficiently. Moreover, we propose a coarse to fine training strategy, which enables fine-grained correspondence prediction. Finally, by matching predicted codes with object surface and using a PnP solver, we estimate the 6DoF pose. Results on the public LM-O and YCB-V datasets show major improvement over the state of the art w.r.t. ADD(-S) metric, even surpassing RGB-D based methods in some cases.

Yan Di, Ruida Zhang, Zhiqiang Lou et al.

While 6D object pose estimation has recently made a huge leap forward, most methods can still only handle a single or a handful of different objects, which limits their applications. To circumvent this problem, category-level object pose estimation has recently been revamped, which aims at predicting the 6D pose as well as the 3D metric size for previously unseen instances from a given set of object classes. This is, however, a much more challenging task due to severe intra-class shape variations. To address this issue, we propose GPV-Pose, a novel framework for robust category-level pose estimation, harnessing geometric insights to enhance the learning of category-level pose-sensitive features. First, we introduce a decoupled confidence-driven rotation representation, which allows geometry-aware recovery of the associated rotation matrix. Second, we propose a novel geometry-guided point-wise voting paradigm for robust retrieval of the 3D object bounding box. Finally, leveraging these different output streams, we can enforce several geometric consistency terms, further increasing performance, especially for non-symmetric categories. GPV-Pose produces superior results to state-of-the-art competitors on common public benchmarks, whilst almost achieving real-time inference speed at 20 FPS.

Stefano Gasperini, Nils Morbitzer, HyunJun Jung et al.

While state-of-the-art monocular depth estimation approaches achieve impressive results in ideal settings, they are highly unreliable under challenging illumination and weather conditions, such as at nighttime or in the presence of rain. In this paper, we uncover these safety-critical issues and tackle them with md4all: a simple and effective solution that works reliably under both adverse and ideal conditions, as well as for different types of learning supervision. We achieve this by exploiting the efficacy of existing methods under perfect settings. Therefore, we provide valid training signals independently of what is in the input. First, we generate a set of complex samples corresponding to the normal training ones. Then, we train the model by guiding its self- or full-supervision by feeding the generated samples and computing the standard losses on the corresponding original images. Doing so enables a single model to recover information across diverse conditions without modifications at inference time. Extensive experiments on two challenging public datasets, namely nuScenes and Oxford RobotCar, demonstrate the effectiveness of our techniques, outperforming prior works by a large margin in both standard and challenging conditions. Source code and data are available at: https://md4all.github.io.

Ruida Zhang, Yan Di, Zhiqiang Lou et al.

Category-level object pose estimation aims to predict the 6D pose as well as the 3D metric size of arbitrary objects from a known set of categories. Recent methods harness shape prior adaptation to map the observed point cloud into the canonical space and apply Umeyama algorithm to recover the pose and size. However, their shape prior integration strategy boosts pose estimation indirectly, which leads to insufficient pose-sensitive feature extraction and slow inference speed. To tackle this problem, in this paper, we propose a novel geometry-guided Residual Object Bounding Box Projection network RBP-Pose that jointly predicts object pose and residual vectors describing the displacements from the shape-prior-indicated object surface projections on the bounding box towards the real surface projections. Such definition of residual vectors is inherently zero-mean and relatively small, and explicitly encapsulates spatial cues of the 3D object for robust and accurate pose regression. We enforce geometry-aware consistency terms to align the predicted pose and residual vectors to further boost performance.

Yamei Chen, Yan Di, Guangyao Zhai et al.

Category-level object pose estimation, aiming to predict the 6D pose and 3D size of objects from known categories, typically struggles with large intra-class shape variation. Existing works utilizing mean shapes often fall short of capturing this variation. To address this issue, we present SecondPose, a novel approach integrating object-specific geometric features with semantic category priors from DINOv2. Leveraging the advantage of DINOv2 in providing SE(3)-consistent semantic features, we hierarchically extract two types of SE(3)-invariant geometric features to further encapsulate local-to-global object-specific information. These geometric features are then point-aligned with DINOv2 features to establish a consistent object representation under SE(3) transformations, facilitating the mapping from camera space to the pre-defined canonical space, thus further enhancing pose estimation. Extensive experiments on NOCS-REAL275 demonstrate that SecondPose achieves a 12.4% leap forward over the state-of-the-art. Moreover, on a more complex dataset HouseCat6D which provides photometrically challenging objects, SecondPose still surpasses other competitors by a large margin.

Ege Özsoy, Evin Pınar Örnek, Ulrich Eck et al.

Surgical procedures are conducted in highly complex operating rooms (OR), comprising different actors, devices, and interactions. To date, only medically trained human experts are capable of understanding all the links and interactions in such a demanding environment. This paper aims to bring the community one step closer to automated, holistic and semantic understanding and modeling of OR domain. Towards this goal, for the first time, we propose using semantic scene graphs (SSG) to describe and summarize the surgical scene. The nodes of the scene graphs represent different actors and objects in the room, such as medical staff, patients, and medical equipment, whereas edges are the relationships between them. To validate the possibilities of the proposed representation, we create the first publicly available 4D surgical SSG dataset, 4D-OR, containing ten simulated total knee replacement surgeries recorded with six RGB-D sensors in a realistic OR simulation center. 4D-OR includes 6734 frames and is richly annotated with SSGs, human and object poses, and clinical roles. We propose an end-to-end neural network-based SSG generation pipeline, with a rate of success of 0.75 macro F1, indeed being able to infer semantic reasoning in the OR. We further demonstrate the representation power of our scene graphs by using it for the problem of clinical role prediction, where we achieve 0.85 macro F1. The code and dataset will be made available upon acceptance.

Guangyao Zhai, Xiaoni Cai, Dianye Huang et al.

Object rearrangement is pivotal in robotic-environment interactions, representing a significant capability in embodied AI. In this paper, we present SG-Bot, a novel rearrangement framework that utilizes a coarse-to-fine scheme with a scene graph as the scene representation. Unlike previous methods that rely on either known goal priors or zero-shot large models, SG-Bot exemplifies lightweight, real-time, and user-controllable characteristics, seamlessly blending the consideration of commonsense knowledge with automatic generation capabilities. SG-Bot employs a three-fold procedure--observation, imagination, and execution--to adeptly address the task. Initially, objects are discerned and extracted from a cluttered scene during the observation. These objects are first coarsely organized and depicted within a scene graph, guided by either commonsense or user-defined criteria. Then, this scene graph subsequently informs a generative model, which forms a fine-grained goal scene considering the shape information from the initial scene and object semantics. Finally, for execution, the initial and envisioned goal scenes are matched to formulate robotic action policies. Experimental results demonstrate that SG-Bot outperforms competitors by a large margin.

Kunyi Li, Michael Niemeyer, Nassir Navab et al.

In recent years, coordinate-based neural implicit representations have shown promising results for the task of Simultaneous Localization and Mapping (SLAM). While achieving impressive performance on small synthetic scenes, these methods often suffer from oversmoothed reconstructions, especially for complex real-world scenes. In this work, we introduce DNS SLAM, a novel neural RGB-D semantic SLAM approach featuring a hybrid representation. Relying only on 2D semantic priors, we propose the first semantic neural SLAM method that trains class-wise scene representations while providing stable camera tracking at the same time. Our method integrates multi-view geometry constraints with image-based feature extraction to improve appearance details and to output color, density, and semantic class information, enabling many downstream applications. To further enable real-time tracking, we introduce a lightweight coarse scene representation which is trained in a self-supervised manner in latent space. Our experimental results achieve state-of-the-art performance on both synthetic data and real-world data tracking while maintaining a commendable operational speed on off-the-shelf hardware. Further, our method outputs class-wise decomposed reconstructions with better texture capturing appearance and geometric details.

Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt et al.

Neural Radiance Field (NeRF) has recently emerged as a powerful representation to synthesize photorealistic novel views. While showing impressive performance, it relies on the availability of dense input views with highly accurate camera poses, thus limiting its application in real-world scenarios. In this work, we introduce Sparse Pose Adjusting Radiance Field (SPARF), to address the challenge of novel-view synthesis given only few wide-baseline input images (as low as 3) with noisy camera poses. Our approach exploits multi-view geometry constraints in order to jointly learn the NeRF and refine the camera poses. By relying on pixel matches extracted between the input views, our multi-view correspondence objective enforces the optimized scene and camera poses to converge to a global and geometrically accurate solution. Our depth consistency loss further encourages the reconstructed scene to be consistent from any viewpoint. Our approach sets a new state of the art in the sparse-view regime on multiple challenging datasets.

Janis Postels, Martin Danelljan, Luc Van Gool et al.

Normalizing Flows (NFs) are flexible explicit generative models that have been shown to accurately model complex real-world data distributions. However, their invertibility constraint imposes limitations on data distributions that reside on lower dimensional manifolds embedded in higher dimensional space. Practically, this shortcoming is often bypassed by adding noise to the data which impacts the quality of the generated samples. In contrast to prior work, we approach this problem by generating samples from the original data distribution given full knowledge about the perturbed distribution and the noise model. To this end, we establish that NFs trained on perturbed data implicitly represent the manifold in regions of maximum likelihood. Then, we propose an optimization objective that recovers the most likely point on the manifold given a sample from the perturbed distribution. Finally, we focus on 3D point clouds for which we utilize the explicit nature of NFs, i.e. surface normals extracted from the gradient of the log-likelihood and the log-likelihood itself, to apply Poisson surface reconstruction to refine generated point sets.

Muhammad Ferjad Naeem, Yongqin Xian, Xiaohua Zhai et al.

Image-Text pretraining on web-scale image caption datasets has become the default recipe for open vocabulary classification and retrieval models thanks to the success of CLIP and its variants. Several works have also used CLIP features for dense prediction tasks and have shown the emergence of open-set abilities. However, the contrastive objective used by these models only focuses on image-text alignment and does not incentivise image feature learning for dense prediction tasks. In this work, we introduce SILC, a novel framework for vision language pretraining. SILC improves image-text contrastive learning with the simple addition of local-to-global correspondence learning by self-distillation. We show that distilling local image features from an exponential moving average (EMA) teacher model significantly improves model performance on dense predictions tasks like detection and segmentation, while also providing improvements on image-level tasks such as classification and retrieval. SILC models sets a new state of the art for zero-shot classification, few shot classification, image and text retrieval, zero-shot segmentation, and open vocabulary segmentation. We further show that SILC features greatly benefit open vocabulary detection, captioning and visual question answering.

Yongzhi Su, Yan Di, Fabian Manhardt et al.

Despite monocular 3D object detection having recently made a significant leap forward thanks to the use of pre-trained depth estimators for pseudo-LiDAR recovery, such two-stage methods typically suffer from overfitting and are incapable of explicitly encapsulating the geometric relation between depth and object bounding box. To overcome this limitation, we instead propose OPA-3D, a single-stage, end-to-end, Occlusion-Aware Pixel-Wise Aggregation network that to jointly estimate dense scene depth with depth-bounding box residuals and object bounding boxes, allowing a two-stream detection of 3D objects, leading to significantly more robust detections. Thereby, the geometry stream denoted as the Geometry Stream, combines visible depth and depth-bounding box residuals to recover the object bounding box via explicit occlusion-aware optimization. In addition, a bounding box based geometry projection scheme is employed in an effort to enhance distance perception. The second stream, named as the Context Stream, directly regresses 3D object location and size. This novel two-stream representation further enables us to enforce cross-stream consistency terms which aligns the outputs of both streams, improving the overall performance. Extensive experiments on the public benchmark demonstrate that OPA-3D outperforms state-of-the-art methods on the main Car category, whilst keeping a real-time inference speed. We plan to release all codes and trained models soon.

Dekai Zhu, Guangyao Zhai, Yan Di et al.

Reliable multi-agent trajectory prediction is crucial for the safe planning and control of autonomous systems. Compared with single-agent cases, the major challenge in simultaneously processing multiple agents lies in modeling complex social interactions caused by various driving intentions and road conditions. Previous methods typically leverage graph-based message propagation or attention mechanism to encapsulate such interactions in the format of marginal probabilistic distributions. However, it is inherently sub-optimal. In this paper, we propose IPCC-TP, a novel relevance-aware module based on Incremental Pearson Correlation Coefficient to improve multi-agent interaction modeling. IPCC-TP learns pairwise joint Gaussian Distributions through the tightly-coupled estimation of the means and covariances according to interactive incremental movements. Our module can be conveniently embedded into existing multi-agent prediction methods to extend original motion distribution decoders. Extensive experiments on nuScenes and Argoverse 2 datasets demonstrate that IPCC-TP improves the performance of baselines by a large margin.

Evin Pınar Örnek, Shristi Mudgal, Johanna Wald et al.

There have been numerous recently proposed methods for monocular depth prediction (MDP) coupled with the equally rapid evolution of benchmarking tools. However, we argue that MDP is currently witnessing benchmark over-fitting and relying on metrics that are only partially helpful to gauge the usefulness of the predictions for 3D applications. This limits the design and development of novel methods that are truly aware of - and improving towards estimating - the 3D structure of the scene rather than optimizing 2D-based distances. In this work, we aim to bring structural awareness to MDP, an inherently 3D task, by exhibiting the limits of evaluation metrics towards assessing the quality of the 3D geometry. We propose a set of metrics well suited to evaluate the 3D geometry of MDP approaches and a novel indoor benchmark, RIO-D3D, crucial for the proposed evaluation methodology. Our benchmark is based on a real-world dataset featuring high-quality rendered depth maps obtained from RGB-D reconstructions. We further demonstrate this to help benchmark the closely-tied task of 3D scene completion.

Yida Wang, David Joseph Tan, Nassir Navab et al.

We propose a novel convolutional operator for the task of point cloud completion. One striking characteristic of our approach is that, conversely to related work it does not require any max-pooling or voxelization operation. Instead, the proposed operator used to learn the point cloud embedding in the encoder extracts permutation-invariant features from the point cloud via a soft-pooling of feature activations, which are able to preserve fine-grained geometric details. These features are then passed on to a decoder architecture. Due to the compression in the encoder, a typical limitation of this type of architectures is that they tend to lose parts of the input shape structure. We propose to overcome this limitation by using skip connections specifically devised for point clouds, where links between corresponding layers in the encoder and the decoder are established. As part of these connections, we introduce a transformation matrix that projects the features from the encoder to the decoder and vice-versa. The quantitative and qualitative results on the task of object completion from partial scans on the ShapeNet dataset show that incorporating our approach achieves state-of-the-art performance in shape completion both at low and high resolutions.

Christina Tsalicoglou, Fabian Manhardt, Alessio Tonioni et al.

The ability to generate highly realistic 2D images from mere text prompts has recently made huge progress in terms of speed and quality, thanks to the advent of image diffusion models. Naturally, the question arises if this can be also achieved in the generation of 3D content from such text prompts. To this end, a new line of methods recently emerged trying to harness diffusion models, trained on 2D images, for supervision of 3D model generation using view dependent prompts. While achieving impressive results, these methods, however, have two major drawbacks. First, rather than commonly used 3D meshes, they instead generate neural radiance fields (NeRFs), making them impractical for most real applications. Second, these approaches tend to produce over-saturated models, giving the output a cartoonish looking effect. Therefore, in this work we propose a novel method for generation of highly realistic-looking 3D meshes. To this end, we extend NeRF to employ an SDF backbone, leading to improved 3D mesh extraction. In addition, we propose a novel way to finetune the mesh texture, removing the effect of high saturation and improving the details of the output 3D mesh.

Chenyangguang Zhang, Guanlong Jiao, Yan Di et al. · tsinghua

Previous works concerning single-view hand-held object reconstruction typically rely on supervision from 3D ground-truth models, which are hard to collect in real world. In contrast, readily accessible hand-object videos offer a promising training data source, but they only give heavily occluded object observations. In this paper, we present a novel synthetic-to-real framework to exploit Multi-view Occlusion-aware supervision from hand-object videos for Hand-held Object reconstruction (MOHO) from a single image, tackling two predominant challenges in such setting: hand-induced occlusion and object's self-occlusion. First, in the synthetic pre-training stage, we render a large-scaled synthetic dataset SOMVideo with hand-object images and multi-view occlusion-free supervisions, adopted to address hand-induced occlusion in both 2D and 3D spaces. Second, in the real-world finetuning stage, MOHO leverages the amodal-mask-weighted geometric supervision to mitigate the unfaithful guidance caused by the hand-occluded supervising views in real world. Moreover, domain-consistent occlusion-aware features are amalgamated in MOHO to resist object's self-occlusion for inferring the complete object shape. Extensive experiments on HO3D and DexYCB datasets demonstrate 2D-supervised MOHO gains superior results against 3D-supervised methods by a large margin.

Artem Savkin, Rachid Ellouze, Nassir Navab et al.

Image synthesis driven by computer graphics achieved recently a remarkable realism, yet synthetic image data generated this way reveals a significant domain gap with respect to real-world data. This is especially true in autonomous driving scenarios, which represent a critical aspect for overcoming utilizing synthetic data for training neural networks. We propose a method based on domain-invariant scene representation to directly synthesize traffic scene imagery without rendering. Specifically, we rely on synthetic scene graphs as our internal representation and introduce an unsupervised neural network architecture for realistic traffic scene synthesis. We enhance synthetic scene graphs with spatial information about the scene and demonstrate the effectiveness of our approach through scene manipulation.

Zhiying Leng, Shun-Cheng Wu, Mahdi Saleh et al.

Reconstructing both objects and hands in 3D from a single RGB image is complex. Existing methods rely on manually defined hand-object constraints in Euclidean space, leading to suboptimal feature learning. Compared with Euclidean space, hyperbolic space better preserves the geometric properties of meshes thanks to its exponentially-growing space distance, which amplifies the differences between the features based on similarity. In this work, we propose the first precise hand-object reconstruction method in hyperbolic space, namely Dynamic Hyperbolic Attention Network (DHANet), which leverages intrinsic properties of hyperbolic space to learn representative features. Our method that projects mesh and image features into a unified hyperbolic space includes two modules, ie. dynamic hyperbolic graph convolution and image-attention hyperbolic graph convolution. With these two modules, our method learns mesh features with rich geometry-image multi-modal information and models better hand-object interaction. Our method provides a promising alternative for fine hand-object reconstruction in hyperbolic space. Extensive experiments on three public datasets demonstrate that our method outperforms most state-of-the-art methods.

Yan Di, Chenyangguang Zhang, Ruida Zhang et al.

In this paper, we propose U-RED, an Unsupervised shape REtrieval and Deformation pipeline that takes an arbitrary object observation as input, typically captured by RGB images or scans, and jointly retrieves and deforms the geometrically similar CAD models from a pre-established database to tightly match the target. Considering existing methods typically fail to handle noisy partial observations, U-RED is designed to address this issue from two aspects. First, since one partial shape may correspond to multiple potential full shapes, the retrieval method must allow such an ambiguous one-to-many relationship. Thereby U-RED learns to project all possible full shapes of a partial target onto the surface of a unit sphere. Then during inference, each sampling on the sphere will yield a feasible retrieval. Second, since real-world partial observations usually contain noticeable noise, a reliable learned metric that measures the similarity between shapes is necessary for stable retrieval. In U-RED, we design a novel point-wise residual-guided metric that allows noise-robust comparison. Extensive experiments on the synthetic datasets PartNet, ComplementMe and the real-world dataset Scan2CAD demonstrate that U-RED surpasses existing state-of-the-art approaches by 47.3%, 16.7% and 31.6% respectively under Chamfer Distance.

Stefano Gasperini, Alvaro Marcos-Ramiro, Michael Schmidt et al.

Panoptic segmentation methods assign a known class to each pixel given in input. Even for state-of-the-art approaches, this inevitably enforces decisions that systematically lead to wrong predictions for objects outside the training categories. However, robustness against out-of-distribution samples and corner cases is crucial in safety-critical settings to avoid dangerous consequences. Since real-world datasets cannot contain enough data points to adequately sample the long tail of the underlying distribution, models must be able to deal with unseen and unknown scenarios as well. Previous methods targeted this by re-identifying already-seen unlabeled objects. In this work, we propose the necessary step to extend segmentation with a new setting which we term holistic segmentation. Holistic segmentation aims to identify and separate objects of unseen, unknown categories into instances without any prior knowledge about them while performing panoptic segmentation of known classes. We tackle this new problem with U3HS, which finds unknowns as highly uncertain regions and clusters their corresponding instance-aware embeddings into individual objects. By doing so, for the first time in panoptic segmentation with unknown objects, our U3HS is trained without unknown categories, reducing assumptions and leaving the settings as unconstrained as in real-life scenarios. Extensive experiments on public data from MS COCO, Cityscapes, and Lost&Found demonstrate the effectiveness of U3HS for this new, challenging, and assumptions-free setting called holistic segmentation. Project page: https://holisticseg.github.io.

Marie-Julie Rakotosaona, Fabian Manhardt, Diego Martin Arroyo et al.

With the introduction of Neural Radiance Fields (NeRFs), novel view synthesis has recently made a big leap forward. At the core, NeRF proposes that each 3D point can emit radiance, allowing to conduct view synthesis using differentiable volumetric rendering. While neural radiance fields can accurately represent 3D scenes for computing the image rendering, 3D meshes are still the main scene representation supported by most computer graphics and simulation pipelines, enabling tasks such as real time rendering and physics-based simulations. Obtaining 3D meshes from neural radiance fields still remains an open challenge since NeRFs are optimized for view synthesis, not enforcing an accurate underlying geometry on the radiance field. We thus propose a novel compact and flexible architecture that enables easy 3D surface reconstruction from any NeRF-driven approach. Upon having trained the radiance field, we distill the volumetric 3D representation into a Signed Surface Approximation Network, allowing easy extraction of the 3D mesh and appearance. Our final 3D mesh is physically accurate and can be rendered in real time on an array of devices.

Bowen Fu, Gu Wang, Chenyangguang Zhang et al. · tsinghua

Reconstructing hand-held objects from a single RGB image is a challenging task in computer vision. In contrast to prior works that utilize deterministic modeling paradigms, we employ a point cloud denoising diffusion model to account for the probabilistic nature of this problem. In the core, we introduce centroid-fixed dual-stream conditional diffusion for monocular hand-held object reconstruction (D-SCo), tackling two predominant challenges. First, to avoid the object centroid from deviating, we utilize a novel hand-constrained centroid fixing paradigm, enhancing the stability of diffusion and reverse processes and the precision of feature projection. Second, we introduce a dual-stream denoiser to semantically and geometrically model hand-object interactions with a novel unified hand-object semantic embedding, enhancing the reconstruction performance of the hand-occluded region of the object. Experiments on the synthetic ObMan dataset and three real-world datasets HO3D, MOW and DexYCB demonstrate that our approach can surpass all other state-of-the-art methods.

Francesco Pinto, Nathalie Rauschmayr, Florian Tramèr et al.

Vision-Language Models (VLMs) have made remarkable progress in document-based Visual Question Answering (i.e., responding to queries about the contents of an input document provided as an image). In this work, we show these models can memorize responses for training samples and regurgitate them even when the relevant visual information has been removed. This includes Personal Identifiable Information (PII) repeated once in the training set, indicating these models could divulge memorised sensitive information and therefore pose a privacy risk. We quantitatively measure the extractability of information in controlled experiments and differentiate between cases where it arises from generalization capabilities or from memorization. We further investigate the factors that influence memorization across multiple state-of-the-art models and propose an effective heuristic countermeasure that empirically prevents the extractability of PII.

Alexander Lehner, Stefano Gasperini, Alvaro Marcos-Ramiro et al.

Since real-world training datasets cannot properly sample the long tail of the underlying data distribution, corner cases and rare out-of-domain samples can severely hinder the performance of state-of-the-art models. This problem becomes even more severe for dense tasks, such as 3D semantic segmentation, where points of non-standard objects can be confidently associated to the wrong class. In this work, we focus on improving the generalization to out-of-domain data. We achieve this by augmenting the training set with adversarial examples. First, we learn a set of vectors that deform the objects in an adversarial fashion. To prevent the adversarial examples from being too far from the existing data distribution, we preserve their plausibility through a series of constraints, ensuring sensor-awareness and shapes smoothness. Then, we perform adversarial augmentation by applying the learned sample-independent vectors to the available objects when training a model. We conduct extensive experiments across a variety of scenarios on data from KITTI, Waymo, and CrashD for 3D object detection, and on data from SemanticKITTI, Waymo, and nuScenes for 3D semantic segmentation. Despite training on a standard single dataset, our approach substantially improves the robustness and generalization of both 3D object detection and 3D semantic segmentation methods to out-of-domain data.

Yan Di, Chenyangguang Zhang, Pengyuan Wang et al.

In this paper, we present a novel shape reconstruction method leveraging diffusion model to generate 3D sparse point cloud for the object captured in a single RGB image. Recent methods typically leverage global embedding or local projection-based features as the condition to guide the diffusion model. However, such strategies fail to consistently align the denoised point cloud with the given image, leading to unstable conditioning and inferior performance. In this paper, we present CCD-3DR, which exploits a novel centered diffusion probabilistic model for consistent local feature conditioning. We constrain the noise and sampled point cloud from the diffusion model into a subspace where the point cloud center remains unchanged during the forward diffusion process and reverse process. The stable point cloud center further serves as an anchor to align each point with its corresponding local projection-based features. Extensive experiments on synthetic benchmark ShapeNet-R2N2 demonstrate that CCD-3DR outperforms all competitors by a large margin, with over 40% improvement. We also provide results on real-world dataset Pix3D to thoroughly demonstrate the potential of CCD-3DR in real-world applications. Codes will be released soon

Ruida Zhang, Yan Di, Fabian Manhardt et al.

Category-level pose estimation is a challenging problem due to intra-class shape variations. Recent methods deform pre-computed shape priors to map the observed point cloud into the normalized object coordinate space and then retrieve the pose via post-processing, i.e., Umeyama's Algorithm. The shortcomings of this two-stage strategy lie in two aspects: 1) The surrogate supervision on the intermediate results can not directly guide the learning of pose, resulting in large pose error after post-processing. 2) The inference speed is limited by the post-processing step. In this paper, to handle these shortcomings, we propose an end-to-end trainable network SSP-Pose for category-level pose estimation, which integrates shape priors into a direct pose regression network. SSP-Pose stacks four individual branches on a shared feature extractor, where two branches are designed to deform and match the prior model with the observed instance, and the other two branches are applied for directly regressing the totally 9 degrees-of-freedom pose and performing symmetry reconstruction and point-wise inlier mask prediction respectively. Consistency loss terms are then naturally exploited to align the outputs of different branches and promote the performance. During inference, only the direct pose regression branch is needed. In this manner, SSP-Pose not only learns category-level pose-sensitive characteristics to boost performance but also keeps a real-time inference speed. Moreover, we utilize the symmetry information of each category to guide the shape prior deformation, and propose a novel symmetry-aware loss to mitigate the matching ambiguity. Extensive experiments on public datasets demonstrate that SSP-Pose produces superior performance compared with competitors with a real-time inference speed at about 25Hz.

Azade Farshad, Yousef Yeganeh, Helisa Dhamo et al.

Graph representation of objects and their relations in a scene, known as a scene graph, provides a precise and discernible interface to manipulate a scene by modifying the nodes or the edges in the graph. Although existing works have shown promising results in modifying the placement and pose of objects, scene manipulation often leads to losing some visual characteristics like the appearance or identity of objects. In this work, we propose DisPositioNet, a model that learns a disentangled representation for each object for the task of image manipulation using scene graphs in a self-supervised manner. Our framework enables the disentanglement of the variational latent embeddings as well as the feature representation in the graph. In addition to producing more realistic images due to the decomposition of features like pose and identity, our method takes advantage of the probabilistic sampling in the intermediate features to generate more diverse images in object replacement or addition tasks. The results of our experiments show that disentangling the feature representations in the latent manifold of the model outperforms the previous works qualitatively and quantitatively on two public benchmarks. Project Page: https://scenegenie.github.io/DispositioNet/

Bharti Munjal, Alessandro Flaborea, Sikandar Amin et al.

Few-shot fine-grained classification and person search appear as distinct tasks and literature has treated them separately. But a closer look unveils important similarities: both tasks target categories that can only be discriminated by specific object details; and the relevant models should generalize to new categories, not seen during training. We propose a novel unified Query-Guided Network (QGN) applicable to both tasks. QGN consists of a Query-guided Siamese-Squeeze-and-Excitation subnetwork which re-weights both the query and gallery features across all network layers, a Query-guided Region Proposal subnetwork for query-specific localisation, and a Query-guided Similarity subnetwork for metric learning. QGN improves on a few recent few-shot fine-grained datasets, outperforming other techniques on CUB by a large margin. QGN also performs competitively on the person search CUHK-SYSU and PRW datasets, where we perform in-depth analysis.

Selim Kuzucu, Alessio Tonioni, Vasile Lup et al.

Large Vision-Language Models (LVLMs) map visual inputs into dense token sequences, imposing a quadratic computational bottleneck for inference. Elastic visual-token compression addresses this by training a single model that can run at multiple visual-token budgets. However, existing approaches struggle under aggressive compression. Spatial-only compression, as in nested pooling, behaves as an imperfect low-pass filter and induces spectral aliasing that obscures fine-grained detail. Query-only compression, as in nested query resampling, replaces explicit grid-aligned tokens with non-local summaries and substantially degrades spatial grounding. To resolve this representational conflict, we introduce PARCEL (Pool-Anchored Resampling with Conditioned Elastic Queries for Efficient Vision-Language Understanding), a visual tokenization architecture that dynamically partitions the labor of feature extraction. PARCEL establishes spatial pool tokens as low-frequency layout anchors and conditions elastic query tokens on these anchors through Pool-Conditioned Query Resampling. This encourages query tokens to focus on complementary visual features rather than redundant spatial mapping. Extensive evaluations across 27 benchmarks show that PARCEL improves the performance-efficiency Pareto frontier, consistently outperforming existing matryoshka baselines across visual-token budgets while preserving the "train once, deploy anywhere" paradigm.

Chenyangguang Zhang, Zhiqiang Lou, Yan Di et al.

Real-time monocular 3D reconstruction is a challenging problem that remains unsolved. Although recent end-to-end methods have demonstrated promising results, tiny structures and geometric boundaries are hardly captured due to their insufficient supervision neglecting spatial details and oversimplified feature fusion ignoring temporal cues. To address the problems, we propose an end-to-end 3D reconstruction network SST, which utilizes Sparse estimated points from visual SLAM system as additional Spatial guidance and fuses Temporal features via a novel cross-modal attention mechanism, achieving more detailed reconstruction results. We propose a Local Spatial-Temporal Fusion module to exploit more informative spatial-temporal cues from multi-view color information and sparse priors, as well a Global Spatial-Temporal Fusion module to refine the local TSDF volumes with the world-frame model from coarse to fine. Extensive experiments on ScanNet and 7-Scenes demonstrate that SST outperforms all state-of-the-art competitors, whilst keeping a high inference speed at 59 FPS, enabling real-world applications with real-time requirements.

Hidenobu Matsuki, Keisuke Tateno, Michael Niemeyer et al.

Neural field-based 3D representations have recently been adopted in many areas including SLAM systems. Current neural SLAM or online mapping systems lead to impressive results in the presence of simple captures, but they rely on a world-centric map representation as only a single neural field model is used. To define such a world-centric representation, accurate and static prior information about the scene, such as its boundaries and initial camera poses, are required. However, in real-time and on-the-fly scene capture applications, this prior knowledge cannot be assumed as fixed or static, since it dynamically changes and it is subject to significant updates based on run-time observations. Particularly in the context of large-scale mapping, significant camera pose drift is inevitable, necessitating the correction via loop closure. To overcome this limitation, we propose NEWTON, a view-centric mapping method that dynamically constructs neural fields based on run-time observation. In contrast to prior works, our method enables camera pose updates using loop closures and scene boundary updates by representing the scene with multiple neural fields, where each is defined in a local coordinate system of a selected keyframe. The experimental results demonstrate the superior performance of our method over existing world-centric neural field-based SLAM systems, in particular for large-scale scenes subject to camera pose updates.

Dario Pavllo, David Joseph Tan, Marie-Julie Rakotosaona et al.

Neural Radiance Fields (NeRF) coupled with GANs represent a promising direction in the area of 3D reconstruction from a single view, owing to their ability to efficiently model arbitrary topologies. Recent work in this area, however, has mostly focused on synthetic datasets where exact ground-truth poses are known, and has overlooked pose estimation, which is important for certain downstream applications such as augmented reality (AR) and robotics. We introduce a principled end-to-end reconstruction framework for natural images, where accurate ground-truth poses are not available. Our approach recovers an SDF-parameterized 3D shape, pose, and appearance from a single image of an object, without exploiting multiple views during training. More specifically, we leverage an unconditional 3D-aware generator, to which we apply a hybrid inversion scheme where a model produces a first guess of the solution which is then refined via optimization. Our framework can de-render an image in as few as 10 steps, enabling its use in practical scenarios. We demonstrate state-of-the-art results on a variety of real and synthetic benchmarks.

Issa Mouawad, Nikolas Brasch, Fabian Manhardt et al.

Monocular 3D object detection continues to attract attention due to the cost benefits and wider availability of RGB cameras. Despite the recent advances and the ability to acquire data at scale, annotation cost and complexity still limit the size of 3D object detection datasets in the supervised settings. Self-supervised methods, on the other hand, aim at training deep networks relying on pretext tasks or various consistency constraints. Moreover, other 3D perception tasks (such as depth estimation) have shown the benefits of temporal priors as a self-supervision signal. In this work, we argue that the temporal consistency on the level of object poses, provides an important supervision signal given the strong prior on physical motion. Specifically, we propose a self-supervised loss which uses this consistency, in addition to render-and-compare losses, to refine noisy pose predictions and derive high-quality pseudo labels. To assess the effectiveness of the proposed method, we finetune a synthetically trained monocular 3D object detection model using the pseudo-labels that we generated on real data. Evaluation on the standard KITTI3D benchmark demonstrates that our method reaches competitive performance compared to other monocular self-supervised and supervised methods.

Sebastian Jung, Artem Zholus, Martin Sundermeyer et al.

Tracking-Any-Point (TAP) models aim to track any point through a video which is a crucial task in AR/XR and robotics applications. The recently introduced TAPNext approach proposes an end-to-end, recurrent transformer architecture to track points frame-by-frame in a purely online fashion -- demonstrating competitive performance at minimal latency. However, we show that TAPNext struggles with longer video sequences and also frequently fails to re-detect query points that reappear after being occluded or leaving the frame. In this work, we present TAPNext++, a model that tracks points in sequences that are orders of magnitude longer while preserving the low memory and compute footprint of the architecture. We train the recurrent video transformer using several data-driven solutions, including training on long 1024-frame sequences enabled by sequence parallelism techniques. We highlight that re-detection performance is a blind spot in the current literature and introduce a new metric, Re-Detection Average Jaccard ($AJ_{RD}$), to explicitly evaluate tracking on re-appearing points. To improve re-detection of points, we introduce tailored geometric augmentations, such as periodic roll that simulates point re-entries, and supervising occluded points. We demonstrate that recurrent transformers can be substantially improved for point tracking and set a new state-of-the-art on multiple benchmarks. Model and code can be found at https://tap-next-plus-plus.github.io.