Eddy Ilg

Jingyun Liang, Yuchen Fan, Xiaoyu Xiang et al.

Video restoration aims at restoring multiple high-quality frames from multiple low-quality frames. Existing video restoration methods generally fall into two extreme cases, i.e., they either restore all frames in parallel or restore the video frame by frame in a recurrent way, which would result in different merits and drawbacks. Typically, the former has the advantage of temporal information fusion. However, it suffers from large model size and intensive memory consumption; the latter has a relatively small model size as it shares parameters across frames; however, it lacks long-range dependency modeling ability and parallelizability. In this paper, we attempt to integrate the advantages of the two cases by proposing a recurrent video restoration transformer, namely RVRT. RVRT processes local neighboring frames in parallel within a globally recurrent framework which can achieve a good trade-off between model size, effectiveness, and efficiency. Specifically, RVRT divides the video into multiple clips and uses the previously inferred clip feature to estimate the subsequent clip feature. Within each clip, different frame features are jointly updated with implicit feature aggregation. Across different clips, the guided deformable attention is designed for clip-to-clip alignment, which predicts multiple relevant locations from the whole inferred clip and aggregates their features by the attention mechanism. Extensive experiments on video super-resolution, deblurring, and denoising show that the proposed RVRT achieves state-of-the-art performance on benchmark datasets with balanced model size, testing memory and runtime.

Leonhard Sommer, Artur Jesslen, Eddy Ilg et al.

Category-level 3D pose estimation is a fundamentally important problem in computer vision and robotics, e.g. for embodied agents or to train 3D generative models. However, so far methods that estimate the category-level object pose require either large amounts of human annotations, CAD models or input from RGB-D sensors. In contrast, we tackle the problem of learning to estimate the category-level 3D pose only from casually taken object-centric videos without human supervision. We propose a two-step pipeline: First, we introduce a multi-view alignment procedure that determines canonical camera poses across videos with a novel and robust cyclic distance formulation for geometric and appearance matching using reconstructed coarse meshes and DINOv2 features. In a second step, the canonical poses and reconstructed meshes enable us to train a model for 3D pose estimation from a single image. In particular, our model learns to estimate dense correspondences between images and a prototypical 3D template by predicting, for each pixel in a 2D image, a feature vector of the corresponding vertex in the template mesh. We demonstrate that our method outperforms all baselines at the unsupervised alignment of object-centric videos by a large margin and provides faithful and robust predictions in-the-wild. Our code and data is available at https://github.com/GenIntel/uns-obj-pose3d.

Christopher Wewer, Eddy Ilg, Bernt Schiele et al.

Existing neural field representations for 3D object reconstruction either (1) utilize object-level representations, but suffer from low-quality details due to conditioning on a global latent code, or (2) are able to perfectly reconstruct the observations, but fail to utilize object-level prior knowledge to infer unobserved regions. We present SimNP, a method to learn category-level self-similarities, which combines the advantages of both worlds by connecting neural point radiance fields with a category-level self-similarity representation. Our contribution is two-fold. (1) We design the first neural point representation on a category level by utilizing the concept of coherent point clouds. The resulting neural point radiance fields store a high level of detail for locally supported object regions. (2) We learn how information is shared between neural points in an unconstrained and unsupervised fashion, which allows to derive unobserved regions of an object during the reconstruction process from given observations. We show that SimNP is able to outperform previous methods in reconstructing symmetric unseen object regions, surpassing methods that build upon category-level or pixel-aligned radiance fields, while providing semantic correspondences between instances

Tom Fischer, Yaoyao Liu, Artur Jesslen et al.

Different from human nature, it is still common practice today for vision tasks to train deep learning models only initially and on fixed datasets. A variety of approaches have recently addressed handling continual data streams. However, extending these methods to manage out-of-distribution (OOD) scenarios has not effectively been investigated. On the other hand, it has recently been shown that non-continual neural mesh models exhibit strong performance in generalizing to such OOD scenarios. To leverage this decisive property in a continual learning setting, we propose incremental neural mesh models that can be extended with new meshes over time. In addition, we present a latent space initialization strategy that enables us to allocate feature space for future unseen classes in advance and a positional regularization term that forces the features of the different classes to consistently stay in respective latent space regions. We demonstrate the effectiveness of our method through extensive experiments on the Pascal3D and ObjectNet3D datasets and show that our approach outperforms the baselines for classification by $2-6\%$ in the in-domain and by $6-50\%$ in the OOD setting. Our work also presents the first incremental learning approach for pose estimation. Our code and model can be found at https://github.com/Fischer-Tom/iNeMo.

Kevin Raj, Christopher Wewer, Raza Yunus et al.

We introduce Spurfies, a novel method for sparse-view surface reconstruction that disentangles appearance and geometry information to utilize local geometry priors trained on synthetic data. Recent research heavily focuses on 3D reconstruction using dense multi-view setups, typically requiring hundreds of images. However, these methods often struggle with few-view scenarios. Existing sparse-view reconstruction techniques often rely on multi-view stereo networks that need to learn joint priors for geometry and appearance from a large amount of data. In contrast, we introduce a neural point representation that disentangles geometry and appearance to train a local geometry prior using a subset of the synthetic ShapeNet dataset only. During inference, we utilize this surface prior as additional constraint for surface and appearance reconstruction from sparse input views via differentiable volume rendering, restricting the space of possible solutions. We validate the effectiveness of our method on the DTU dataset and demonstrate that it outperforms previous state of the art by 35% in surface quality while achieving competitive novel view synthesis quality. Moreover, in contrast to previous works, our method can be applied to larger, unbounded scenes, such as Mip-NeRF 360.

Jonas Belouadi, Eddy Ilg, Margret Keuper et al.

Automatically synthesizing figures from text captions is a compelling capability. However, achieving high geometric precision and editability requires representing figures as graphics programs in languages like TikZ, and aligned training data (i.e., graphics programs with captions) remains scarce. Meanwhile, large amounts of unaligned graphics programs and captioned raster images are more readily available. We reconcile these disparate data sources by presenting TikZero, which decouples graphics program generation from text understanding by using image representations as an intermediary bridge. It enables independent training on graphics programs and captioned images and allows for zero-shot text-guided graphics program synthesis during inference. We show that our method substantially outperforms baselines that can only operate with caption-aligned graphics programs. Furthermore, when leveraging caption-aligned graphics programs as a complementary training signal, TikZero matches or exceeds the performance of much larger models, including commercial systems like GPT-4o. Our code, datasets, and select models are publicly available.

Agniv Sharma, Xianghui Xie, Tom Fischer et al.

Modeling and generating 3D human-object interactions from text is crucial for applications in AR, XR, and gaming. Existing approaches often rely on score distillation from text-to-image models, but their results suffer from the Janus problem and do not follow text prompts faithfully due to the scarcity of high-quality interaction data. We introduce Hoi3DGen, a framework that generates high-quality textured meshes of human-object interaction that follow the input interaction descriptions precisely. We first curate realistic and high-quality interaction data leveraging multimodal large language models, and then create a full text-to-3D pipeline, which achieves orders-of-magnitude improvements in interaction fidelity. Our method surpasses baselines by 4-15x in text consistency and 3-7x in 3D model quality, exhibiting strong generalization to diverse categories and interaction types, while maintaining high-quality 3D generation.

Tom Fischer, Xiaojie Zhang, Eddy Ilg

Recognizing objects in images is a fundamental problem in computer vision. Although detecting objects in 2D images is common, many applications require determining their pose in 3D space. Traditional category-level methods rely on RGB-D inputs, which may not always be available, or employ two-stage approaches that use separate models and representations for detection and pose estimation. For the first time, we introduce a unified model that integrates detection and pose estimation into a single framework for RGB images by leveraging neural mesh models with learned features and multi-model RANSAC. Our approach achieves state-of-the-art results for RGB category-level pose estimation on REAL275, improving on the current state-of-the-art by 22.9% averaged across all scale-agnostic metrics. Finally, we demonstrate that our unified method exhibits greater robustness compared to single-stage baselines. Our code and models are available at https://github.com/Fischer-Tom/unified-detection-and-pose-estimation.

Osama Makansi, Eddy Ilg, Özgün Cicek et al.

Future prediction is a fundamental principle of intelligence that helps plan actions and avoid possible dangers. As the future is uncertain to a large extent, modeling the uncertainty and multimodality of the future states is of great relevance. Existing approaches are rather limited in this regard and mostly yield a single hypothesis of the future or, at the best, strongly constrained mixture components that suffer from instabilities in training and mode collapse. In this work, we present an approach that involves the prediction of several samples of the future with a winner-takes-all loss and iterative grouping of samples to multiple modes. Moreover, we discuss how to evaluate predicted multimodal distributions, including the common real scenario, where only a single sample from the ground-truth distribution is available for evaluation. We show on synthetic and real data that the proposed approach triggers good estimates of multimodal distributions and avoids mode collapse. Source code is available at $\href{https://github.com/lmb-freiburg/Multimodal-Future-Prediction}{\text{this https URL.}}$

Christopher Wewer, Kevin Raj, Eddy Ilg et al.

We present latentSplat, a method to predict semantic Gaussians in a 3D latent space that can be splatted and decoded by a light-weight generative 2D architecture. Existing methods for generalizable 3D reconstruction either do not scale to large scenes and resolutions, or are limited to interpolation of close input views. latentSplat combines the strengths of regression-based and generative approaches while being trained purely on readily available real video data. The core of our method are variational 3D Gaussians, a representation that efficiently encodes varying uncertainty within a latent space consisting of 3D feature Gaussians. From these Gaussians, specific instances can be sampled and rendered via efficient splatting and a fast, generative decoder. We show that latentSplat outperforms previous works in reconstruction quality and generalization, while being fast and scalable to high-resolution data.

Raza Yunus, Jan Eric Lenssen, Michael Niemeyer et al.

Reconstructing models of the real world, including 3D geometry, appearance, and motion of real scenes, is essential for computer graphics and computer vision. It enables the synthesizing of photorealistic novel views, useful for the movie industry and AR/VR applications. It also facilitates the content creation necessary in computer games and AR/VR by avoiding laborious manual design processes. Further, such models are fundamental for intelligent computing systems that need to interpret real-world scenes and actions to act and interact safely with the human world. Notably, the world surrounding us is dynamic, and reconstructing models of dynamic, non-rigidly moving scenes is a severely underconstrained and challenging problem. This state-of-the-art report (STAR) offers the reader a comprehensive summary of state-of-the-art techniques with monocular and multi-view inputs such as data from RGB and RGB-D sensors, among others, conveying an understanding of different approaches, their potential applications, and promising further research directions. The report covers 3D reconstruction of general non-rigid scenes and further addresses the techniques for scene decomposition, editing and controlling, and generalizable and generative modeling. More specifically, we first review the common and fundamental concepts necessary to understand and navigate the field and then discuss the state-of-the-art techniques by reviewing recent approaches that use traditional and machine-learning-based neural representations, including a discussion on the newly enabled applications. The STAR is concluded with a discussion of the remaining limitations and open challenges.

Cameron Braunstein, Mariya Toneva, Eddy Ilg

Latent diffusion models such as Stable Diffusion achieve state-of-the-art results on text-to-image generation tasks. However, the extent to which these models have a semantic understanding of the images they generate is not well understood. In this work, we investigate whether the internal representations used by these models during text-to-image generation contain semantic information that is meaningful to humans. To do so, we perform probing on Stable Diffusion with simple regression layers that predict semantic attributes for objects and evaluate these predictions against human annotations. Surprisingly, we find that this success can actually be attributed to the text encoding occurring in CLIP rather than the reverse diffusion process. We demonstrate that groups of specific semantic attributes have markedly different decoding accuracy than the average, and are thus represented to different degrees. Finally, we show that attributes become more difficult to disambiguate from one another during the inverse diffusion process, further demonstrating the strongest semantic representation of object attributes in CLIP. We conclude that the separately trained CLIP vision-language model is what determines the human-like semantic representation, and that the diffusion process instead takes the role of a visual decoder.

Philipp Schröppel, Christopher Wewer, Jan Eric Lenssen et al.

Controllable generation of 3D assets is important for many practical applications like content creation in movies, games and engineering, as well as in AR/VR. Recently, diffusion models have shown remarkable results in generation quality of 3D objects. However, none of the existing models enable disentangled generation to control the shape and appearance separately. For the first time, we present a suitable representation for 3D diffusion models to enable such disentanglement by introducing a hybrid point cloud and neural radiance field approach. We model a diffusion process over point positions jointly with a high-dimensional feature space for a local density and radiance decoder. While the point positions represent the coarse shape of the object, the point features allow modeling the geometry and appearance details. This disentanglement enables us to sample both independently and therefore to control both separately. Our approach sets a new state of the art in generation compared to previous disentanglement-capable methods by reduced FID scores of 30-90% and is on-par with other non disentanglement-capable state-of-the art methods.

Jonas Kälble, Sascha Wirges, Maxim Tatarchenko et al.

Automated driving fundamentally requires knowledge about the surrounding geometry of the scene. Modern approaches use only captured images to predict occupancy maps that represent the geometry. Training these approaches requires accurate data that may be acquired with the help of LiDAR scanners. We show that the techniques used for current benchmarks and training datasets to convert LiDAR scans into occupancy grid maps yield very low quality, and subsequently present a novel approach using evidence theory that yields more accurate reconstructions. We demonstrate that these are superior by a large margin, both qualitatively and quantitatively, and that we additionally obtain meaningful uncertainty estimates. When converting the occupancy maps back to depth estimates and comparing them with the raw LiDAR measurements, our method yields a MAE improvement of 30% to 52% on nuScenes and 53% on Waymo over other occupancy ground-truth data. Finally, we use the improved occupancy maps to train a state-of-the-art occupancy prediction method and demonstrate that it improves the MAE by 25% on nuScenes.

Sohaib Zahid, Viktor Rudnev, Eddy Ilg et al.

Novel view synthesis techniques predominantly utilize RGB cameras, inheriting their limitations such as the need for sufficient lighting, susceptibility to motion blur, and restricted dynamic range. In contrast, event cameras are significantly more resilient to these limitations but have been less explored in this domain, particularly in large-scale settings. Current methodologies primarily focus on front-facing or object-oriented (360-degree view) scenarios. For the first time, we introduce 3D Gaussians for event-based novel view synthesis. Our method reconstructs large and unbounded scenes with high visual quality. We contribute the first real and synthetic event datasets tailored for this setting. Our method demonstrates superior novel view synthesis and consistently outperforms the baseline EventNeRF by a margin of 11-25% in PSNR (dB) while being orders of magnitude faster in reconstruction and rendering.

Michael Goesele, Daniel Andersen, Yujia Chen et al.

In recent years smart glasses technology has rapidly advanced, opening up entirely new areas for mobile computing. We expect future smart glasses will need to be all-day wearable, adopting a small form factor to meet the requirements of volume, weight, fashionability and social acceptability, which puts significant constraints on the space of possible solutions. Additional challenges arise due to the fact that smart glasses are worn in arbitrary environments while their wearer moves and performs everyday activities. In this paper, we systematically analyze the space of imaging from smart glasses and derive several fundamental limits that govern this imaging domain. We discuss the impact of these limits on achievable image quality and camera module size -- comparing in particular to related devices such as mobile phones. We then propose a novel distributed imaging approach that allows to minimize the size of the individual camera modules when compared to a standard monolithic camera design. Finally, we demonstrate the properties of this novel approach in a series of experiments using synthetic data as well as images captured with two different prototype implementations.

Noor Ahmed, Cameron Braunstein, Steffen Eger et al.

Recent Multi-Modal Large Language Models (MLLMs) have demonstrated strong capabilities in learning joint representations from text and images. However, their spatial reasoning remains limited. We introduce 3DFroMLLM, a novel framework that enables the generation of 3D object prototypes directly from MLLMs, including geometry and part labels. Our pipeline is agentic, comprising a designer, coder, and visual inspector operating in a refinement loop. Notably, our approach requires no additional training data or detailed user instructions. Building on prior work in 2D generation, we demonstrate that rendered images produced by our framework can be effectively used for image classification pretraining tasks and outperforms previous methods by 15%. As a compelling real-world use case, we show that the generated prototypes can be leveraged to improve fine-grained vision-language models by using the rendered, part-labeled prototypes to fine-tune CLIP for part segmentation and achieving a 55% accuracy improvement without relying on any additional human-labeled data.

Cameron Braunstein, Hevra Petekkaya, Jan Eric Lenssen et al.

We introduce SLayR, Scene Layout Generation with Rectified flow, a novel transformer-based model for text-to-layout generation which can then be paired with existing layout-to-image models to produce images. SLayR addresses a domain in which current text-to-image pipelines struggle: generating scene layouts that are of significant variety and plausibility, when the given prompt is ambiguous and does not provide constraints on the scene. SLayR surpasses existing baselines including LLMs in unconstrained generation, and can generate layouts from an open caption set. To accurately evaluate the layout generation, we introduce a new benchmark suite, including numerical metrics and a carefully designed repeatable human-evaluation procedure that assesses the plausibility and variety of generated images. We show that our method sets a new state of the art for achieving both at the same time, while being at least 3x times smaller in the number of parameters.

Cameron Braunstein, Eddy Ilg, Vladislav Golyanik

Quantum visual computing is advancing rapidly. This paper presents a new formulation for stereo matching with nonlinear regularizers and spatial pyramids on quantum annealers as a maximum a posteriori inference problem that minimizes the energy of a Markov Random Field. Our approach is hybrid (i.e., quantum-classical) and is compatible with modern D-Wave quantum annealers, i.e., it includes a quadratic unconstrained binary optimization (QUBO) objective. Previous quantum annealing techniques for stereo matching are limited to using linear regularizers, and thus, they do not exploit the fundamental advantages of the quantum computing paradigm in solving combinatorial optimization problems. In contrast, our method utilizes the full potential of quantum annealing for stereo matching, as nonlinear regularizers create optimization problems which are NP-hard. On the Middlebury benchmark, we achieve an improved root mean squared accuracy over the previous state of the art in quantum stereo matching of 2% and 22.5% when using different solvers.

Tom Fischer, Pascal Peter, Joachim Weickert et al.

Deep learning has revolutionized the field of computer vision by introducing large scale neural networks with millions of parameters. Training these networks requires massive datasets and leads to intransparent models that can fail to generalize. At the other extreme, models designed from partial differential equations (PDEs) embed specialized domain knowledge into mathematical equations and usually rely on few manually chosen hyperparameters. This makes them transparent by construction and if designed and calibrated carefully, they can generalize well to unseen scenarios. In this paper, we show how to bring model- and data-driven approaches together by combining the explicit PDE-based approaches with convolutional neural networks to obtain the best of both worlds. We illustrate a joint architecture for the task of inpainting optical flow fields and show that the combination of model- and data-driven modeling leads to an effective architecture. Our model outperforms both fully explicit and fully data-driven baselines in terms of reconstruction quality, robustness and amount of required training data. Averaging the endpoint error across different mask densities, our method outperforms the explicit baselines by 11-27%, the GAN baseline by 47% and the Probabilisitic Diffusion baseline by 42%. With that, our method sets a new state of the art for inpainting of optical flow fields from random masks.

Samir Aroudj, Steven Lovegrove, Eddy Ilg et al.

We propose a novel explicit dense 3D reconstruction approach that processes a set of images of a scene with sensor poses and calibrations and estimates a photo-real digital model. One of the key innovations is that the underlying volumetric representation is completely explicit in contrast to neural network-based (implicit) alternatives. We encode scenes explicitly using clear and understandable mappings of optimization variables to scene geometry and their outgoing surface radiance. We represent them using hierarchical volumetric fields stored in a sparse voxel octree. Robustly reconstructing such a volumetric scene model with millions of unknown variables from registered scene images only is a highly non-convex and complex optimization problem. To this end, we employ stochastic gradient descent (Adam) which is steered by an inverse differentiable renderer. We demonstrate that our method can reconstruct models of high quality that are comparable to state-of-the-art implicit methods. Importantly, we do not use a sequential reconstruction pipeline where individual steps suffer from incomplete or unreliable information from previous stages, but start our optimizations from uniformed initial solutions with scene geometry and radiance that is far off from the ground truth. We show that our method is general and practical. It does not require a highly controlled lab setup for capturing, but allows for reconstructing scenes with a vast variety of objects, including challenging ones, such as outdoor plants or furry toys. Finally, our reconstructed scene models are versatile thanks to their explicit design. They can be edited interactively which is computationally too costly for implicit alternatives.

Tony Ng, Hyo Jin Kim, Vincent Lee et al.

In the light of recent analyses on privacy-concerning scene revelation from visual descriptors, we develop descriptors that conceal the input image content. In particular, we propose an adversarial learning framework for training visual descriptors that prevent image reconstruction, while maintaining the matching accuracy. We let a feature encoding network and image reconstruction network compete with each other, such that the feature encoder tries to impede the image reconstruction with its generated descriptors, while the reconstructor tries to recover the input image from the descriptors. The experimental results demonstrate that the visual descriptors obtained with our method significantly deteriorate the image reconstruction quality with minimal impact on correspondence matching and camera localization performance.

Deeksha Dangwal, Vincent T. Lee, Hyo Jin Kim et al.

As autonomous driving and augmented reality evolve, a practical concern is data privacy. In particular, these applications rely on localization based on user images. The widely adopted technology uses local feature descriptors, which are derived from the images and it was long thought that they could not be reverted back. However, recent work has demonstrated that under certain conditions reverse engineering attacks are possible and allow an adversary to reconstruct RGB images. This poses a potential risk to user privacy. We take this a step further and model potential adversaries using a privacy threat model. Subsequently, we show under controlled conditions a reverse engineering attack on sparse feature maps and analyze the vulnerability of popular descriptors including FREAK, SIFT and SOSNet. Finally, we evaluate potential mitigation techniques that select a subset of descriptors to carefully balance privacy reconstruction risk while preserving image matching accuracy; our results show that similar accuracy can be obtained when revealing less information.

Sungyong Baik, Hyo Jin Kim, Tianwei Shen et al.

We tackle the problem of visual localization under changing conditions, such as time of day, weather, and seasons. Recent learned local features based on deep neural networks have shown superior performance over classical hand-crafted local features. However, in a real-world scenario, there often exists a large domain gap between training and target images, which can significantly degrade the localization accuracy. While existing methods utilize a large amount of data to tackle the problem, we present a novel and practical approach, where only a few examples are needed to reduce the domain gap. In particular, we propose a few-shot domain adaptation framework for learned local features that deals with varying conditions in visual localization. The experimental results demonstrate the superior performance over baselines, while using a scarce number of training examples from the target domain.

Wenxin Liu, David Caruso, Eddy Ilg et al.

In this work we propose a tightly-coupled Extended Kalman Filter framework for IMU-only state estimation. Strap-down IMU measurements provide relative state estimates based on IMU kinematic motion model. However the integration of measurements is sensitive to sensor bias and noise, causing significant drift within seconds. Recent research by Yan et al. (RoNIN) and Chen et al. (IONet) showed the capability of using trained neural networks to obtain accurate 2D displacement estimates from segments of IMU data and obtained good position estimates from concatenating them. This paper demonstrates a network that regresses 3D displacement estimates and its uncertainty, giving us the ability to tightly fuse the relative state measurement into a stochastic cloning EKF to solve for pose, velocity and sensor biases. We show that our network, trained with pedestrian data from a headset, can produce statistically consistent measurement and uncertainty to be used as the update step in the filter, and the tightly-coupled system outperforms velocity integration approaches in position estimates, and AHRS attitude filter in orientation estimates.

Rohan Chabra, Jan Eric Lenssen, Eddy Ilg et al.

Efficiently reconstructing complex and intricate surfaces at scale is a long-standing goal in machine perception. To address this problem we introduce Deep Local Shapes (DeepLS), a deep shape representation that enables encoding and reconstruction of high-quality 3D shapes without prohibitive memory requirements. DeepLS replaces the dense volumetric signed distance function (SDF) representation used in traditional surface reconstruction systems with a set of locally learned continuous SDFs defined by a neural network, inspired by recent work such as DeepSDF. Unlike DeepSDF, which represents an object-level SDF with a neural network and a single latent code, we store a grid of independent latent codes, each responsible for storing information about surfaces in a small local neighborhood. This decomposition of scenes into local shapes simplifies the prior distribution that the network must learn, and also enables efficient inference. We demonstrate the effectiveness and generalization power of DeepLS by showing object shape encoding and reconstructions of full scenes, where DeepLS delivers high compression, accuracy, and local shape completion.

Osama Makansi, Eddy Ilg, Thomas Brox

Recent work has shown that convolutional neural networks (CNNs) can be used to estimate optical flow with high quality and fast runtime. This makes them preferable for real-world applications. However, such networks require very large training datasets. Engineering the training data is difficult and/or laborious. This paper shows how to augment a network trained on an existing synthetic dataset with large amounts of additional unlabelled data. In particular, we introduce a selection mechanism to assemble from multiple estimates a joint optical flow field, which outperforms that of all input methods. The latter can be used as proxy-ground-truth to train a network on real-world data and to adapt it to specific domains of interest. Our experimental results show that the performance of networks improves considerably, both, in cross-domain and in domain-specific scenarios. As a consequence, we obtain state-of-the-art results on the KITTI benchmarks.

Eddy Ilg, Tonmoy Saikia, Margret Keuper et al.

Occlusions play an important role in disparity and optical flow estimation, since matching costs are not available in occluded areas and occlusions indicate depth or motion boundaries. Moreover, occlusions are relevant for motion segmentation and scene flow estimation. In this paper, we present an efficient learning-based approach to estimate occlusion areas jointly with disparities or optical flow. The estimated occlusions and motion boundaries clearly improve over the state-of-the-art. Moreover, we present networks with state-of-the-art performance on the popular KITTI benchmark and good generic performance. Making use of the estimated occlusions, we also show improved results on motion segmentation and scene flow estimation.

Eddy Ilg, Özgün Çiçek, Silvio Galesso et al.

Optical flow estimation can be formulated as an end-to-end supervised learning problem, which yields estimates with a superior accuracy-runtime tradeoff compared to alternative methodology. In this paper, we make such networks estimate their local uncertainty about the correctness of their prediction, which is vital information when building decisions on top of the estimations. For the first time we compare several strategies and techniques to estimate uncertainty in a large-scale computer vision task like optical flow estimation. Moreover, we introduce a new network architecture utilizing the Winner-Takes-All loss and show that this can provide complementary hypotheses and uncertainty estimates efficiently with a single forward pass and without the need for sampling or ensembles. Finally, we demonstrate the quality of the different uncertainty estimates, which is clearly above previous confidence measures on optical flow and allows for interactive frame rates.

Nikolaus Mayer, Eddy Ilg, Philipp Fischer et al.

The finding that very large networks can be trained efficiently and reliably has led to a paradigm shift in computer vision from engineered solutions to learning formulations. As a result, the research challenge shifts from devising algorithms to creating suitable and abundant training data for supervised learning. How to efficiently create such training data? The dominant data acquisition method in visual recognition is based on web data and manual annotation. Yet, for many computer vision problems, such as stereo or optical flow estimation, this approach is not feasible because humans cannot manually enter a pixel-accurate flow field. In this paper, we promote the use of synthetically generated data for the purpose of training deep networks on such tasks.We suggest multiple ways to generate such data and evaluate the influence of dataset properties on the performance and generalization properties of the resulting networks. We also demonstrate the benefit of learning schedules that use different types of data at selected stages of the training process.

Osama Makansi, Eddy Ilg, Thomas Brox

Learning approaches have shown great success in the task of super-resolving an image given a low resolution input. Video super-resolution aims for exploiting additionally the information from multiple images. Typically, the images are related via optical flow and consecutive image warping. In this paper, we provide an end-to-end video super-resolution network that, in contrast to previous works, includes the estimation of optical flow in the overall network architecture. We analyze the usage of optical flow for video super-resolution and find that common off-the-shelf image warping does not allow video super-resolution to benefit much from optical flow. We rather propose an operation for motion compensation that performs warping from low to high resolution directly. We show that with this network configuration, video super-resolution can benefit from optical flow and we obtain state-of-the-art results on the popular test sets. We also show that the processing of whole images rather than independent patches is responsible for a large increase in accuracy.

Anna Khoreva, Rodrigo Benenson, Eddy Ilg et al.

Convolutional networks reach top quality in pixel-level video object segmentation but require a large amount of training data (1k~100k) to deliver such results. We propose a new training strategy which achieves state-of-the-art results across three evaluation datasets while using 20x~1000x less annotated data than competing methods. Our approach is suitable for both single and multiple object segmentation. Instead of using large training sets hoping to generalize across domains, we generate in-domain training data using the provided annotation on the first frame of each video to synthesize ("lucid dream") plausible future video frames. In-domain per-video training data allows us to train high quality appearance- and motion-based models, as well as tune the post-processing stage. This approach allows to reach competitive results even when training from only a single annotated frame, without ImageNet pre-training. Our results indicate that using a larger training set is not automatically better, and that for the video object segmentation task a smaller training set that is closer to the target domain is more effective. This changes the mindset regarding how many training samples and general "objectness" knowledge are required for the video object segmentation task.

Benjamin Ummenhofer, Huizhong Zhou, Jonas Uhrig et al.

In this paper we formulate structure from motion as a learning problem. We train a convolutional network end-to-end to compute depth and camera motion from successive, unconstrained image pairs. The architecture is composed of multiple stacked encoder-decoder networks, the core part being an iterative network that is able to improve its own predictions. The network estimates not only depth and motion, but additionally surface normals, optical flow between the images and confidence of the matching. A crucial component of the approach is a training loss based on spatial relative differences. Compared to traditional two-frame structure from motion methods, results are more accurate and more robust. In contrast to the popular depth-from-single-image networks, DeMoN learns the concept of matching and, thus, better generalizes to structures not seen during training.

Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia et al.

The FlowNet demonstrated that optical flow estimation can be cast as a learning problem. However, the state of the art with regard to the quality of the flow has still been defined by traditional methods. Particularly on small displacements and real-world data, FlowNet cannot compete with variational methods. In this paper, we advance the concept of end-to-end learning of optical flow and make it work really well. The large improvements in quality and speed are caused by three major contributions: first, we focus on the training data and show that the schedule of presenting data during training is very important. Second, we develop a stacked architecture that includes warping of the second image with intermediate optical flow. Third, we elaborate on small displacements by introducing a sub-network specializing on small motions. FlowNet 2.0 is only marginally slower than the original FlowNet but decreases the estimation error by more than 50%. It performs on par with state-of-the-art methods, while running at interactive frame rates. Moreover, we present faster variants that allow optical flow computation at up to 140fps with accuracy matching the original FlowNet.

Nikolaus Mayer, Eddy Ilg, Philip Häusser et al.

Recent work has shown that optical flow estimation can be formulated as a supervised learning task and can be successfully solved with convolutional networks. Training of the so-called FlowNet was enabled by a large synthetically generated dataset. The present paper extends the concept of optical flow estimation via convolutional networks to disparity and scene flow estimation. To this end, we propose three synthetic stereo video datasets with sufficient realism, variation, and size to successfully train large networks. Our datasets are the first large-scale datasets to enable training and evaluating scene flow methods. Besides the datasets, we present a convolutional network for real-time disparity estimation that provides state-of-the-art results. By combining a flow and disparity estimation network and training it jointly, we demonstrate the first scene flow estimation with a convolutional network.

Philipp Fischer, Alexey Dosovitskiy, Eddy Ilg et al.

Convolutional neural networks (CNNs) have recently been very successful in a variety of computer vision tasks, especially on those linked to recognition. Optical flow estimation has not been among the tasks where CNNs were successful. In this paper we construct appropriate CNNs which are capable of solving the optical flow estimation problem as a supervised learning task. We propose and compare two architectures: a generic architecture and another one including a layer that correlates feature vectors at different image locations. Since existing ground truth data sets are not sufficiently large to train a CNN, we generate a synthetic Flying Chairs dataset. We show that networks trained on this unrealistic data still generalize very well to existing datasets such as Sintel and KITTI, achieving competitive accuracy at frame rates of 5 to 10 fps.