3D reconstruction methods such as Neural Radiance Fields (NeRFs) excel at rendering photorealistic novel views of complex scenes. However, recovering a high-quality NeRF typically requires tens to hundreds of input images, resulting in a time-consuming capture process. We present ReconFusion to reconstruct real-world scenes using only a few photos. Our approach leverages a diffusion prior for novel view synthesis, trained on synthetic and multiview datasets, which regularizes a NeRF-based 3D reconstruction pipeline at novel camera poses beyond those captured by the set of input images. Our method synthesizes realistic geometry and texture in underconstrained regions while preserving the appearance of observed regions. We perform an extensive evaluation across various real-world datasets, including forward-facing and 360-degree scenes, demonstrating significant performance improvements over previous few-view NeRF reconstruction approaches.
Neural Radiance Fields (NeRF) can be optimized to obtain high-fidelity 3D scene reconstructions of objects and large-scale scenes. However, NeRFs require accurate camera parameters as input -- inaccurate camera parameters result in blurry renderings. Extrinsic and intrinsic camera parameters are usually estimated using Structure-from-Motion (SfM) methods as a pre-processing step to NeRF, but these techniques rarely yield perfect estimates. Thus, prior works have proposed jointly optimizing camera parameters alongside a NeRF, but these methods are prone to local minima in challenging settings. In this work, we analyze how different camera parameterizations affect this joint optimization problem, and observe that standard parameterizations exhibit large differences in magnitude with respect to small perturbations, which can lead to an ill-conditioned optimization problem. We propose using a proxy problem to compute a whitening transform that eliminates the correlation between camera parameters and normalizes their effects, and we propose to use this transform as a preconditioner for the camera parameters during joint optimization. Our preconditioned camera optimization significantly improves reconstruction quality on scenes from the Mip-NeRF 360 dataset: we reduce error rates (RMSE) by 67% compared to state-of-the-art NeRF approaches that do not optimize for cameras like Zip-NeRF, and by 29% relative to state-of-the-art joint optimization approaches using the camera parameterization of SCNeRF. Our approach is easy to implement, does not significantly increase runtime, can be applied to a wide variety of camera parameterizations, and can straightforwardly be incorporated into other NeRF-like models.
Traditional approaches for learning 3D object categories have been predominantly trained and evaluated on synthetic datasets due to the unavailability of real 3D-annotated category-centric data. Our main goal is to facilitate advances in this field by collecting real-world data in a magnitude similar to the existing synthetic counterparts. The principal contribution of this work is thus a large-scale dataset, called Common Objects in 3D, with real multi-view images of object categories annotated with camera poses and ground truth 3D point clouds. The dataset contains a total of 1.5 million frames from nearly 19,000 videos capturing objects from 50 MS-COCO categories and, as such, it is significantly larger than alternatives both in terms of the number of categories and objects. We exploit this new dataset to conduct one of the first large-scale "in-the-wild" evaluations of several new-view-synthesis and category-centric 3D reconstruction methods. Finally, we contribute NerFormer - a novel neural rendering method that leverages the powerful Transformer to reconstruct an object given a small number of its views. The CO3D dataset is available at https://github.com/facebookresearch/co3d .
Our goal is to learn a deep network that, given a small number of images of an object of a given category, reconstructs it in 3D. While several recent works have obtained analogous results using synthetic data or assuming the availability of 2D primitives such as keypoints, we are interested in working with challenging real data and with no manual annotations. We thus focus on learning a model from multiple views of a large collection of object instances. We contribute with a new large dataset of object centric videos suitable for training and benchmarking this class of models. We show that existing techniques leveraging meshes, voxels, or implicit surfaces, which work well for reconstructing isolated objects, fail on this challenging data. Finally, we propose a new neural network design, called warp-conditioned ray embedding (WCR), which significantly improves reconstruction while obtaining a detailed implicit representation of the object surface and texture, also compensating for the noise in the initial SfM reconstruction that bootstrapped the learning process. Our evaluation demonstrates performance improvements over several deep monocular reconstruction baselines on existing benchmarks and on our novel dataset.
We learn a latent space for easy capture, semantic editing, consistent interpolation, and efficient reproduction of visual material appearance. When users provide a photo of a stationary natural material captured under flash light illumination, it is converted in milliseconds into a latent material code. In a second step, conditioned on the material code, our method, again in milliseconds, produces an infinite and diverse spatial field of BRDF model parameters (diffuse albedo, specular albedo, roughness, normals) that allows rendering in complex scenes and illuminations, matching the appearance of the input picture. Technically, we jointly embed all flash images into a latent space using a convolutional encoder, and -- conditioned on these latent codes -- convert random spatial fields into fields of BRDF parameters using a convolutional neural network (CNN). We condition these BRDF parameters to match the visual characteristics (statistics and spectra of visual features) of the input under matching light. A user study confirms that the semantics of the latent material space agree with user expectations and compares our approach favorably to previous work.
We propose a generative model of 2D and 3D natural textures with diversity, visual fidelity and at high computational efficiency. This is enabled by a family of methods that extend ideas from classic stochastic procedural texturing (Perlin noise) to learned, deep, non-linearities. The key idea is a hard-coded, tunable and differentiable step that feeds multiple transformed random 2D or 3D fields into an MLP that can be sampled over infinite domains. Our model encodes all exemplars from a diverse set of textures without a need to be re-trained for each exemplar. Applications include texture interpolation, and learning 3D textures from 2D exemplars.
We develop PlatonicGAN to discover 3D structure of an object class from an unstructured collection of 2D images. The key idea is to learn a deep neural network that generates 3D shapes that are never objectionable to a discriminator looking only at its 2D projections, i.e. renderings of the generated volumes. Using such a 2D instead of a 3D discriminator allows tapping into massive 2D image collections instead of relying on much smaller 3D data sets. To establish constraints between 2D image observation and their 3D interpretation we suggest a family of rendering layers that are effectively back-propagatable. This family includes visual hull, absorption-only (akin to x-ray), and emission-absorption (that can resolve occlusion if multiple 3D points project to the same 2D pixel). These layers are studied both on synthetic and real data in an application to reconstruct of 3D shape from 2D images.
The comprehensive representation and understanding of the driving environment is crucial to improve the safety and reliability of autonomous vehicles. In this paper, we present a new approach to establish an environment model containing a segmentation between static and dynamic background and parametric modeled objects with shape, position and orientation. Multiple laser scanners are fused into a dynamic occupancy grid map resulting in a 360{\deg} perception of the environment. A single-stage deep convolutional neural network is combined with a recurrent neural network, which takes a time series of the occupancy grid map as input and tracks cell states and its corresponding object hypotheses. The labels for training are created unsupervised with an automatic label generation algorithm. The proposed methods are evaluated in real-world experiments in complex inner city scenarios using the aforementioned 360{\deg} laser perception. The results show a better object detection accuracy in comparison with our old approach as well as an AUC score of 0.946 for the dynamic and static segmentation. Furthermore, we gain an improved detection for occluded objects and a more consistent size estimation due to the usage of time series as input and the memory about previous states introduced by the recurrent neural network.
We tackle the problem of object detection and pose estimation in a shared space downtown environment. For perception multiple laser scanners with 360{\deg} coverage were fused in a dynamic occupancy grid map (DOGMa). A single-stage deep convolutional neural network is trained to provide object hypotheses comprising of shape, position, orientation and an existence score from a single input DOGMa. Furthermore, an algorithm for offline object extraction was developed to automatically label several hours of training data. The algorithm is based on a two-pass trajectory extraction, forward and backward in time. Typical for engineered algorithms, the automatic label generation suffers from misdetections, which makes hard negative mining impractical. Therefore, we propose a loss function counteracting the high imbalance between mostly static background and extremely rare dynamic grid cells. Experiments indicate, that the trained network has good generalization capabilities since it detects objects occasionally lost by the label algorithm. Evaluation reaches an average precision (AP) of 75.9%