Recovering the 3D geometry of a scene from a sparse set of uncalibrated images is a long-standing problem in computer vision. While recent learning-based approaches such as DUSt3R and MASt3R have demonstrated impressive results by directly predicting dense scene geometry, they are primarily trained on outdoor scenes with static environments and struggle to handle human-centric scenarios. In this work, we introduce HAMSt3R, an extension of MASt3R for joint human and scene 3D reconstruction from sparse, uncalibrated multi-view images. First, we exploit DUNE, a strong image encoder obtained by distilling, among others, the encoders from MASt3R and from a state-of-the-art Human Mesh Recovery (HMR) model, multi-HMR, for a better understanding of scene geometry and human bodies. Our method then incorporates additional network heads to segment people, estimate dense correspondences via DensePose, and predict depth in human-centric environments, enabling a more comprehensive 3D reconstruction. By leveraging the outputs of our different heads, HAMSt3R produces a dense point map enriched with human semantic information in 3D. Unlike existing methods that rely on complex optimization pipelines, our approach is fully feed-forward and efficient, making it suitable for real-world applications. We evaluate our model on EgoHumans and EgoExo4D, two challenging benchmarks con taining diverse human-centric scenarios. Additionally, we validate its generalization to traditional multi-view stereo and multi-view pose regression tasks. Our results demonstrate that our method can reconstruct humans effectively while preserving strong performance in general 3D reconstruction tasks, bridging the gap between human and scene understanding in 3D vision.

Human perception and understanding is a major domain of computer vision which, like many other vision subdo-mains recently, stands to gain from the use of large models pre-trained on large datasets. We hypothesize that the most common pre-training strategy of relying on general purpose, object-centric image datasets such as ImageNet, is limited by an important domain shift. On the other hand, collecting domain-specific ground truth such as 2D or 3D labels does not scale well. Therefore, we propose a pre-training approach based on self-supervised learning that works on human-centric data using only images. Our method uses pairs of images of humans: the first is partially masked and the model is trained to reconstruct the masked parts given the visible ones and a second image. It relies on both stereoscopic (cross-view) pairs, and tempo-ral (cross-pose) pairs taken from videos, in order to learn priors about 3D as well as human motion. We pre-train a model for body-centric tasks and one for hand-centric tasks. With a generic transformer architecture, these models outperform existing self-supervised pre-training methods on a wide set of human-centric downstream tasks, and obtain state-of-the-art performance for instance when fine-tuning for model-based and model-free human mesh recovery.

Existing neural human rendering methods struggle with a single image input due to the lack of information in in-visible areas and the depth ambiguity of pixels in visible areas. In this regard, we propose Monocular Neural Human Renderer (MonoNHR), a novel approach that renders robust free-viewpoint images of an arbitrary human given only a single image. MonoNHR is the first method that (i) renders human subjects never seen during training in a monocular setup, and (ii) is trained in a weakly-supervised manner without geometry supervision. First, we propose to disentangle 3D geometry and texture features and to condition the texture inference on the 3D geometry features. Second, we introduce a Mesh Inpainter module that inpaints the occluded parts exploiting human structural priors such as symmetry. Experiments on ZJU-MoCap, AIST and HUMBI datasets show that our approach significantly outperforms the recent methods adapted to the monocular case.

We examine the problem of mesh denoising, which consists of removing noise from corrupted 3D meshes while preserving existing geometric features. Most mesh denoising methods require a lot of mesh-specific parameter fine-tuning, to account for specific features and noise types. In recent years, data-driven methods have demonstrated their robustness and effectiveness with respect to noise and feature properties on a wide variety of geometry and image problems. Most existing mesh denoising methods still use hand-crafted features, and locally denoise facets rather than examine the mesh globally. In this work, we propose the use of a fully end-to-end learning strategy based on graph convolutions, where meaningful features are learned directly by our network. It operates on a graph of facets, directly on the existing topology of the mesh, without resampling, and follows a multi-scale design to extract geometric features at different resolution levels. Similar to most recent pipelines, given a noisy mesh, we first denoise face normals with our novel approach, then update vertex positions accordingly. Our method performs significantly better than the current state-of-the-art learning-based methods. Additionally, we show that it can be trained on noisy data, without explicit correspondence between noisy and ground-truth facets. We also propose a multi-scale denoising strategy, better suited to correct noise with a low spatial frequency.

In this paper we report on the representation of appearance information in the context of 3D multi-view shape modeling. Most applications in image based 3D modeling resort to texture maps, a 2D mapping of shape color information into image files. Despite their unquestionable merits, in particular the ability to apply standard image tools, including compression, image textures still suffer from limitations that result from the 2D mapping of information that originally belongs to a 3D structure. This is especially true with 2D texture atlases, a generic 2D mapping for 3D mesh models that introduces discontinuities in the texture space and plagues many 3D appearance algorithms. Moreover, the per-triangle texel density of 2D image textures cannot be individually adjusted to the corresponding pixel observation density without a global change in the atlas mapping function. To address these issues, we propose a new appearance representation for image-based 3D shape modeling, which stores appearance information directly on 3D meshes, rather than a texture atlas. We show this representation to allow for input-adaptive sampling and compression support. Our experiments demonstrate that it outperforms traditional image textures, in multi-view reconstruction contexts, with better visual quality and memory footprint, which makes it a suitable tool when dealing with large amounts of data as with dynamic scene 3D models.