We propose a method for generating fly-through videos of a scene, from a single image and a given camera trajectory. We build upon an image-to-video latent diffusion model [5]. We condition its UNet [25] denoiser on the camera trajectory, using four techniques. (1) We condition UNet's temporal blocks on raw camera extrinsics, similar to MotionCtrl [36]. (2) We use images containing camera ray parameters, similar to CameraCtrl [14]. (3) We re-project the initial image to subsequent frames and condition on the resulting video. (4) We introduce a global 3D representation using 2D ⇔ 3D transformers [32], which implicitly conditions on the camera poses. We combine all conditions in a ContolNet-style [42] architecture. We then propose a metric that evaluates overall video quality and the ability to preserve details with view changes, which we use to analyze the trade-offs of individual and combined conditions. Finally, we identify an optimal combination of conditions. We calibrate camera positions in our datasets for scale consistency across scenes, and we train our scene exploration model, CamCtrl3D, demonstrating state-of-the-art results.

We propose a method for annotating videos of complex multi-object scenes with a globally-consistent 3D representation of the objects. We annotate each object with a CAD model from a database, and place it in the 3D coordinate frame of the scene with a 9-DoF pose transformation. Our method is semi-automatic and works on commonly-available RGB videos, without requiring a depth sensor. Many steps are performed automatically, and the tasks performed by humans are simple, well-specified, and require only limited reasoning in 3D. This makes them feasible for crowd-sourcing and has allowed us to construct a large-scale dataset by annotating real-estate videos from YouTube. Our dataset CAD-Estate offers 101k instances of 12k unique CAD models placed in the 3D representations of 20k videos. In comparison to Scan2CAD, the largest existing dataset with CAD model annotations on real scenes, CAD-Estate has 7× more instances and 4× more unique CAD models. We showcase the benefits of pre-training a Mask2CAD model on CAD-Estate for the task of automatic 3D object reconstruction and pose estimation, demonstrating that it leads to performance improvements on the popular Scan2CAD benchmark. The dataset is available at https://github.com/google-research/cad-estate.

Recent advances in neural reconstruction enable high-quality 3D object reconstruction from casually captured image collections. Current techniques mostly analyze their progress on relatively simple image collections where SfM techniques can provide ground-truth (GT) camera poses. We note that SfM techniques tend to fail on in-the-wild image collections such as image search results with varying backgrounds and illuminations. To enable systematic research progress on 3D reconstruction from casual image captures, we propose `NAVI': a new dataset of category-agnostic image collections of objects with high-quality 3D scans along with per-image 2D-3D alignments providing near-perfect GT camera parameters. These 2D-3D alignments allow us to extract accurate derivative annotations such as dense pixel correspondences, depth and segmentation maps. We demonstrate the use of NAVI image collections on different problem settings and show that NAVI enables more thorough evaluations that were not possible with existing datasets. We believe NAVI is beneficial for systematic research progress on 3D reconstruction and correspondence estimation.

Indoor rooms are among the most common use cases in 3D scene understanding. Current state-of-the-art methods for this task are driven by large annotated datasets. Room layouts are especially important, consisting of structural elements in 3D, such as wall, floor, and ceiling. However, they are difficult to annotate, especially on pure RGB video. We propose a novel method to produce generic 3D room layouts just from 2D segmentation masks, which are easy to annotate for humans. Based on these 2D annotations, we automatically reconstruct 3D plane equations for the structural elements and their spatial extent in the scene, and connect adjacent elements at the appropriate contact edges. We annotate and publicly release 2246 3D room layouts on the RealEstate10k dataset, containing YouTube videos. We demonstrate the high quality of these 3D layouts annotations with extensive experiments.

We address the task of aligning CAD models to a video sequence of a complex scene containing multiple objects. Our method can process arbitrary videos and fully automatically recover the 9 DoF pose for each object appearing in it, thus aligning them in a common 3D coordinate frame. The core idea of our method is to integrate neural network predictions from individual frames with a temporally global, multi-view constraint optimization formulation. This integration process resolves the scale and depth ambiguities in the per-frame predictions, and generally improves the estimate of all pose parameters. By leveraging multi-view constraints, our method also resolves occlusions and handles objects that are out of view in individual frames, thus reconstructing all objects into a single globally consistent CAD representation of the scene. In comparison to the state-of-the-art single-frame method Mask2CAD that we build on, we achieve substantial improvements on the Scan2CAD dataset (from 11.6% to 30.7% class average accuracy).

Flow-based generative models have highly desirable properties like exact log-likelihood evaluation and exact latent-variable inference, however they are still in their infancy and have not received as much attention as alternative generative models. In this paper, we introduce C-Flow, a novel conditioning scheme that brings normalizing flows to an entirely new scenario with great possibilities for multimodal data modeling. C-Flow is based on a parallel sequence of invertible mappings in which a source flow guides the target flow at every step, enabling fine-grained control over the generation process. We also devise a new strategy to model unordered 3D point clouds that, in combination with the conditioning scheme, makes it possible to address 3D reconstruction from a single image and its inverse problem of rendering an image given a point cloud. We demonstrate our conditioning method to be very adaptable, being also applicable to image manipulation, style transfer and multi-modal image-to-image mapping in a diversity of domains, including RGB images, segmentation maps and edge masks.

We present Open Images V4, a dataset of 9.2M images with unified annotations for image classification, object detection and visual relationship detection. The images have a Creative Commons Attribution license that allows to share and adapt the material, and they have been collected from Flickr without a predefined list of class names or tags, leading to natural class statistics and avoiding an initial design bias. Open Images V4 offers large scale across several dimensions: 30.1M image-level labels for 19.8k concepts, 15.4M bounding boxes for 600 object classes, and 375k visual relationship annotations involving 57 classes. For object detection in particular, we provide 15x more bounding boxes than the next largest datasets (15.4M boxes on 1.9M images). The images often show complex scenes with several objects (8 annotated objects per image on average). We annotated visual relationships between them, which support visual relationship detection, an emerging task that requires structured reasoning. We provide in-depth comprehensive statistics about the dataset, we validate the quality of the annotations, we study how the performance of several modern models evolves with increasing amounts of training data, and we demonstrate two applications made possible by having unified annotations of multiple types coexisting in the same images. We hope that the scale, quality, and variety of Open Images V4 will foster further research and innovation even beyond the areas of image classification, object detection, and visual relationship detection.

Manually annotating object segmentation masks is very time consuming. Interactive object segmentation methods offer a more efficient alternative where a human annotator and a machine segmentation model collaborate. In this paper we make several contributions to interactive segmentation: (1) we systematically explore in simulation the design space of deep interactive segmentation models and report new insights and caveats; (2) we execute a large-scale annotation campaign with real human annotators, producing masks for 2.5M instances on the OpenImages dataset. We released this data publicly, forming the largest existing dataset for instance segmentation. Moreover, by re-annotating part of the COCO dataset, we show that we can produce instance masks 3x faster than traditional polygon drawing tools while also providing better quality. (3) We present a technique for automatically estimating the quality of the produced masks which exploits indirect signals from the annotation process.

We propose to revisit knowledge transfer for training object detectors on target classes from weakly supervised training images, helped by a set of source classes with bounding-box annotations. We present a unified knowledge transfer framework based on training a single neural network multi-class object detector over all source classes, organized in a semantic hierarchy. This generates proposals with scores at multiple levels in the hierarchy, which we use to explore knowledge transfer over a broad range of generality, ranging from class-specific (bycicle to motorbike) to class-generic (objectness to any class). Experiments on the 200 object classes in the ILSVRC 2013 detection dataset show that our technique (1) leads to much better performance on the target classes (70.3% CorLoc, 36.9% mAP) than a weakly supervised baseline which uses manually engineered objectness [11] (50.5% CorLoc, 25.4% mAP). (2) delivers target object detectors reaching 80% of the mAP of their fully supervised counterparts. (3) outperforms the best reported transfer learning results on this dataset (+41% CorLoc and +3% mAP over [18, 46], +16.2% mAP over [32]). Moreover, we also carry out several across-dataset knowledge transfer experiments [27, 24, 35] and find that (4) our technique outperforms the weakly supervised baseline in all dataset pairs by 1.5 Ã- - 1.9Ã-, establishing its general applicability.

Abstract Bidirectional path tracing (BDPT) with Multiple Importance Sampling is one of the most versatile unbiased rendering algorithms today. BDPT repeatedly generates sub‐paths from the eye and the lights, which are connected for each pixel and then discarded. Unfortunately, many such bidirectional connections turn out to have low contribution to the solution. Our key observation is that we can importance sample connections to an eye sub‐path by considering multiple light sub‐paths at once and creating connections probabilistically. We do this by storing light paths, and estimating probability mass functions of the discrete set of possible connections to all light paths. This has two key advantages: we efficiently create connections with low variance by Monte Carlo sampling, and we reuse light paths across different eye paths. We also introduce a caching scheme by deriving an approximation to sub‐path contribution which avoids high‐dimensional path distance computations. Our approach builds on caching methods developed in the different context of VPLs. Our Probabilistic Connections for Bidirectional Path Tracing approach raises a major challenge, since reuse results in high variance due to correlation between paths. We analyze the problem of path correlation and derive a conservative upper bound of the variance, with computationally tractable sample weights. We present results of our method which shows significant improvement over previous unbiased global illumination methods, and evaluate our algorithmic choices.

We introduce a method to compute intrinsic images for a multiview set of outdoor photos with cast shadows, taken under the same lighting. We use an automatic 3D reconstruction from these photos and the sun direction as input and decompose each image into reflectance and shading layers, despite the inaccuracies and missing data of the 3D model. Our approach is based on two key ideas. First, we progressively improve the accuracy of the parameters of our image formation model by performing iterative estimation and combining 3D lighting simulation with 2D image optimization methods. Second, we use the image formation model to express reflectance as a function of discrete visibility values for shadow and light, which allows to introduce a robust visibility classifier for pairs of points in a scene. This classifier is used for shadow labeling, allowing to compute high-quality reflectance and shading layers. Our multiview intrinsic decomposition is of sufficient quality to allow relighting of the input images. We create shadow-caster geometry which preserves shadow silhouettes and, using the intrinsic layers, we can perform multiview relighting with moving cast shadows. We present results on several multiview datasets, and show how it is now possible to perform image-based rendering with changing illumination conditions.

Abstract Ray tracing has become a viable alternative to rasterization for interactive applications and also forms the basis of most global illumination methods. However, even today's fastest ray‐tracers offer only a tight budget of rays per pixel per frame. Rendering performance can be improved by increasing this budget, or by developing methods that use it more efficiently. In this paper we propose a global visibility caching algorithm that reduces the number of shadow rays required for shading to a fraction of less than 2% in some cases. We quantize the visibility function's domain while ensuring a minimal degradation of the final image quality. To control the introduced error, we adapt the quantization locally, accounting for variations in geometry, sampling densities on both endpoints of the visibility queries, and the light signal itself. Compared to previous approaches for approximating visibility, e.g. shadow mapping, our method has several advantages: (1) it allows caching of arbitrary visibility queries between surface points and is thus applicable to all ray tracing based methods; (2) the approximation error is uniform over the entire image and can be bounded by a user‐specified parameter; (3) the cache is created on‐the‐fly and does not waste any resources on queries that will never be used. We demonstrate the benefits of our method on Whitted‐style ray tracing combined with instant radiosity, as well as an integration with bidirectional path tracing.

Vector graphics represent images with compact, editable and scalable primitives. Skillful vector artists employ these primitives to produce vivid depictions of material appearance and lighting. However, such stylized imagery often requires building complex multi-layered combinations of colored fills and gradient meshes. We facilitate this task by introducing vector shade trees that bring to vector graphics the flexibility of modular shading representations as known in the 3D rendering community. In contrast to traditional shade trees that combine pixel and vertex shaders, our shade nodes encapsulate the creation and blending of vector primitives that vector artists routinely use. We propose a set of basic shade nodes that we design to respect the traditional guidelines on material depiction described in drawing books and tutorials. We integrate our representation as an Adobe Illustrator plug-in that allows even inexperienced users to take a line drawing, apply a few clicks and obtain a fully colored illustration. More experienced artists can easily refine the illustration, adding more details and visual features, while using all the vector drawing tools they are already familiar with. We demonstrate the power of our representation by quickly generating illustrations of complex objects and materials.

Abstract Realistic rendering requires computing the global illumination in the scene, and Monte Carlo integration is the best‐known method for doing that. The key to good performance is to carefully select the costly integration samples, which is usually achieved via importance sampling. Unfortunately, visibility is difficult to factor into the importance distribution, which can greatly increase variance in highly occluded scenes with complex illumination. In this paper, we present importance caching – a novel approach that selects those samples with a distribution that includes visibility, while maintaining efficiency by exploiting illumination smoothness. At a sparse set of locations in the scene, we construct and cache several types of probability distributions with respect to a set of virtual point lights (VPLs), which notably include visibility. Each distribution type is optimized for a specific lighting condition. For every shading point, we then borrow the distributions from nearby cached locations and use them for VPL sampling, avoiding additional bias. A novel multiple importance sampling framework finally combines the many estimators. In highly occluded scenes, where visibility is a major source of variance in the incident radiance, our approach can reduce variance by more than an order of magnitude. Even in such complex scenes we can obtain accurate and low noise previews with full global illumination in a couple of seconds on a single mid‐range CPU.

Abstract Significant advances have been achieved for realtime ray tracing recently, but realtime performance for complex scenes still requires large computational resources not yet available from the CPUs in standard PCs. Incidentally, most of these PCs also contain modern GPUs that do offer much larger raw compute power. However, limitations in the programming and memory model have so far kept the performance of GPU ray tracers well below that of their CPU counterparts. In this paper we present a novel packet ray traversal implementation that completely eliminates the need for maintaining a stack during kd‐tree traversal and that reduces the number of traversal steps per ray. While CPUs benefit moderately from the stackless approach, it improves GPU performance significantly. We achieve a peak performance of over 16 million rays per second for reasonably complex scenes, including complex shading and secondary rays. Several examples show that with this new technique GPUs can actually outperform equivalent CPU based ray tracers.