Self-supervised visual foundation models produce powerful embeddings that achieve remarkable performance on a wide range of downstream tasks. However, unlike vision-language models such as CLIP [67], self-supervised visual features are not readily aligned with language, hindering their adoption in open-vocabulary tasks. Our method, named dino.txt, unlocks this new ability for DINOv2 [63], a widely used self-supervised visual encoder. We build upon the LiT training strategy [97], which trains a text encoder to align with a frozen vision model but leads to unsatisfactory results on dense tasks. We propose several key ingredients to improve performance on both global and dense tasks, such as concatenating the [CLS] token with the patch average to train the alignment and curating data using both text and image modalities. With these, we successfully train a CLIP-like model with only a fraction of the computational cost compared to CLIP while achieving state-of-the-art results in zero-shot classification and open-vocabulary semantic segmentation.

We present NeuroMorph, a new neural network architecture that takes as input two 3D shapes and produces in one go, i.e. in a single feed forward pass, a smooth interpolation and point-to-point correspondences between them. The interpolation, expressed as a deformation field, changes the pose of the source shape to resemble the target, but leaves the object identity unchanged. NeuroMorph uses an elegant architecture combining graph convolutions with global feature pooling to extract local features. During training, the model is incentivized to create realistic deformations by approximating geodesics on the underlying shape space manifold. This strong geometric prior allows to train our model end-to-end and in a fully unsupervised manner without requiring any manual correspondence annotations. NeuroMorph works well for a large variety of input shapes, including non-isometric pairs from different object categories. It obtains state-of-the-art results for both shape correspondence and interpolation tasks, matching or surpassing the performance of recent unsupervised and supervised methods on multiple benchmarks.

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 avail-ability 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. For additional material please visit: https://henzler.github.io/publication/unsupervised_videos/.

We tackle the problem of learning the geometry of multiple categories of deformable objects jointly. Recent work has shown that it is possible to learn a unified dense pose predictor for several categories of related objects. However, training such models requires to initialize inter-category correspondences by hand. This is suboptimal and the resulting models fail to maintain correct correspondences as individual categories are learned. In this paper, we show that improved correspondences can be learned automatically as a natural byproduct of learning category-specific dense pose predictors. To do this, we express correspondences between different categories and between images and categories using a unified embedding. Then, we use the latter to enforce two constraints: symmetric inter-category cycle consistency and a new asymmetric image-to-category cycle consistency. Without any manual annotations for the intercategory correspondences, we obtain state-of-the-art alignment results, outperforming dedicated methods for matching 3D shapes. Moreover, the new model is also better at the task of dense pose prediction than prior work.

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.

We tackle the problem of monocular 3D reconstruction of articulated objects like humans and animals. We contribute DensePose 3D, a method that can learn such reconstructions in a weakly supervised fashion from 2D image annotations only. This is in stark contrast with previous deformable reconstruction methods that use parametric models such as SMPL pre-trained on a large dataset of 3D object scans. Because it does not require 3D scans, DensePose 3D can be used for learning a wide range of articulated categories such as different animal species. The method learns, in an end-to-end fashion, a soft partition of a given category-specific 3D template mesh into rigid parts together with a monocular reconstruction network that predicts the part motions such that they reproject correctly onto 2D DensePose-like surface annotations of the object. The decomposition of the object into parts is regularized by expressing part assignments as a combination of the smooth eigenfunctions of the Laplace-Beltrami operator. We show significant improvements compared to state-of-the-art nonrigid structure-from-motion baselines on both synthetic and real data on categories of humans and animals.

In this work, we focus on the task of learning and representing dense correspondences in deformable object categories. While this problem has been considered before, solutions so far have been rather ad-hoc for specific object types (i.e., humans), often with significant manual work involved. However, scaling the geometry understanding to all objects in nature requires more automated approaches that can also express correspondences between related, but geometrically different objects. To this end, we propose a new, learnable image-based representation of dense correspondences. Our model predicts, for each pixel in a 2D image, an embedding vector of the corresponding vertex in the object mesh, therefore establishing dense correspondences between image pixels and 3D object geometry. We demonstrate that the proposed approach performs on par or better than the state-of-the-art methods for dense pose estimation for humans, while being conceptually simpler. We also collect a new in-the-wild dataset of dense correspondences for animal classes and demonstrate that our framework scales naturally to the new deformable object categories.

Since the initial comparison of Seitz et al., the accuracy of dense multiview stereovision methods has been increasing steadily. A number of limitations, however, make most of these methods not suitable to outdoor scenes taken under uncontrolled imaging conditions. The present work consists of a complete dense multiview stereo pipeline which circumvents these limitations, being able to handle large-scale scenes without sacrificing accuracy. Highly detailed reconstructions are produced within very reasonable time thanks to two key stages in our pipeline: a minimum s-t cut optimization over an adaptive domain that robustly and efficiently filters a quasidense point cloud from outliers and reconstructs an initial surface by integrating visibility constraints, followed by a mesh-based variational refinement that captures small details, smartly handling photo-consistency, regularization, and adaptive resolution. The pipeline has been tested over a wide range of scenes: from classic compact objects taken in a laboratory setting, to outdoor architectural scenes, landscapes, and cultural heritage sites. The accuracy of its reconstructions has also been measured on the dense multiview benchmark proposed by Strecha et al., showing the results to compare more than favorably with the current state-of-the-art methods.

In this paper, we present a novel method, the first to date to our knowledge, which is capable of directly and automatically producing a concise and idealized 3D representation from unstructured point data of complex cluttered real-world scenes, with a high level of noise and a significant proportion of outliers, such as those obtained from passive stereo. Our algorithm can digest millions of input points into an optimized lightweight watertight polygonal mesh free of self-intersection, that preserves the structural components of the scene at a user-defined scale, and completes missing scene parts in a plausible manner. To achieve this, our algorithm incorporates priors on urban and architectural scenes, notably the prevalence of vertical structures and orthogonal intersections. A major contribution of our work is an adaptive decomposition of 3D space induced by planar primitives, namely a polyhedral cell complex. We experimentally validate our approach on several challenging noisy point clouds of urban and architectural scenes.

Abstract We describe a robust but simple algorithm to reconstruct a surface from a set of merged range scans. Our key contribution is the formulation of the surface reconstruction problem as an energy minimisation problem that explicitly models the scanning process. The adaptivity of the Delaunay triangulation is exploited by restricting the energy to inside/outside labelings of Delaunay tetrahedra. Our energy measures both the output surface quality and how well the surface agrees with soft visibility constraints. Such energy is shown to perfectly fit into the minimum s − t cuts optimisation framework, allowing fast computation of a globally optimal tetrahedra labeling, while avoiding the “shrinking bias” that usually plagues graph cuts methods. The behaviour of our method confronted to noise, undersampling and outliers is evaluated on several data sets and compared with other methods through different experiments: its strong robustness would make our method practical not only for reconstruction from range data but also from typically more difficult dense point clouds, resulting for instance from stereo image matching. Our effective modeling of the surface acquisition inverse problem, along with the unique combination of Delaunay triangulation and minimum s − t cuts, makes the computational requirements of the algorithm scale well with respect to the size of the input point cloud.

We present a novel method to reconstruct the 3D shape of a scene from several calibrated images. Our motivation is that most existing multi-view stereovision approaches require some knowledge of the scene extent and often even of its approximate geometry (e.g. visual hull). This makes these approaches mainly suited to compact objects admitting a tight enclosing box, imaged on a simple or a known background. In contrast, our approach focuses on large-scale cluttered scenes under uncontrolled imaging conditions. It first generates a quasi-dense 3D point cloud of the scene by matching keypoints across images in a lenient manner, thus possibly retaining many false matches. Then it builds an adaptive tetrahedral decomposition of space by computing the 3D Delaunay triangulation of the 3D point set. Finally, it reconstructs the scene by labeling Delaunay tetrahedra as empty or occupied, thus generating a triangular mesh of the scene. A globally optimal label assignment, as regards photo-consistency of the output mesh and compatibility with the visibility of keypoints in input images, is efficiently found as a minimum cut solution in a graph.