Volume rendering-based 3D reconstruction from multi-view images has gained popularity in recent years, largely due to the success of neural radiance fields (NeRF). A number of methods have been developed that build upon NeRF and use neural volume rendering to learn signed distance fields (SDFs) for reconstructing 3D models. However, SDF-based methods cannot represent non-watertight models and, therefore, cannot capture open boundaries. This paper proposes a new algorithm for learning an accurate unsigned distance field (UDF) from multi-view images, which is specifically designed for reconstructing non-watertight, textureless models. The proposed method, called NeUDF, addresses the limitations of existing UDF-based methods by introducing a simple and approximately unbiased and occlusion-aware density function. In addition, a smooth and differentiable UDF representation is presented to make the learning process easier and more efficient. Experiments on both texture-rich and textureless models demonstrate the robustness and effectiveness of the proposed approach, making it a promising solution for reconstructing challenging 3D models from multi-view images.
相關內容
Coordinate-based implicit neural networks, or neural fields, have emerged as useful representations of shape and appearance in 3D computer vision. Despite advances, however, it remains challenging to build neural fields for categories of objects without datasets like ShapeNet that provide "canonicalized" object instances that are consistently aligned for their 3D position and orientation (pose). We present Canonical Field Network (CaFi-Net), a self-supervised method to canonicalize the 3D pose of instances from an object category represented as neural fields, specifically neural radiance fields (NeRFs). CaFi-Net directly learns from continuous and noisy radiance fields using a Siamese network architecture that is designed to extract equivariant field features for category-level canonicalization. During inference, our method takes pre-trained neural radiance fields of novel object instances at arbitrary 3D pose and estimates a canonical field with consistent 3D pose across the entire category. Extensive experiments on a new dataset of 1300 NeRF models across 13 object categories show that our method matches or exceeds the performance of 3D point cloud-based methods.
A fully automated object reconstruction pipeline is crucial for digital content creation. While the area of 3D reconstruction has witnessed profound developments, the removal of background to obtain a clean object model still relies on different forms of manual labor, such as bounding box labeling, mask annotations, and mesh manipulations. In this paper, we propose a novel framework named AutoRecon for the automated discovery and reconstruction of an object from multi-view images. We demonstrate that foreground objects can be robustly located and segmented from SfM point clouds by leveraging self-supervised 2D vision transformer features. Then, we reconstruct decomposed neural scene representations with dense supervision provided by the decomposed point clouds, resulting in accurate object reconstruction and segmentation. Experiments on the DTU, BlendedMVS and CO3D-V2 datasets demonstrate the effectiveness and robustness of AutoRecon.
In this work, we propose an inverse rendering model that estimates 3D shape, spatially-varying reflectance, homogeneous subsurface scattering parameters, and an environment illumination jointly from only a pair of captured images of a translucent object. In order to solve the ambiguity problem of inverse rendering, we use a physically-based renderer and a neural renderer for scene reconstruction and material editing. Because two renderers are differentiable, we can compute a reconstruction loss to assist parameter estimation. To enhance the supervision of the proposed neural renderer, we also propose an augmented loss. In addition, we use a flash and no-flash image pair as the input. To supervise the training, we constructed a large-scale synthetic dataset of translucent objects, which consists of 117K scenes. Qualitative and quantitative results on both synthetic and real-world datasets demonstrated the effectiveness of the proposed model.
We consider a category-level perception problem, where one is given 2D or 3D sensor data picturing an object of a given category (e.g., a car), and has to reconstruct the 3D pose and shape of the object despite intra-class variability (i.e., different car models have different shapes). We consider an active shape model, where -for an object category- we are given a library of potential CAD models describing objects in that category, and we adopt a standard formulation where pose and shape are estimated from 2D or 3D keypoints via non-convex optimization. Our first contribution is to develop PACE3D* and PACE2D*, the first certifiably optimal solvers for pose and shape estimation using 3D and 2D keypoints, respectively. Both solvers rely on the design of tight (i.e., exact) semidefinite relaxations. Our second contribution is to develop outlier-robust versions of both solvers, named PACE3D# and PACE2D#. Towards this goal, we propose ROBIN, a general graph-theoretic framework to prune outliers, which uses compatibility hypergraphs to model measurements' compatibility. We show that in category-level perception problems these hypergraphs can be built from the winding orders of the keypoints (in 2D) or their convex hulls (in 3D), and many outliers can be filtered out via maximum hyperclique computation. The last contribution is an extensive experimental evaluation. Besides providing an ablation study on simulated datasets and on the PASCAL3D+ dataset, we combine our solver with a deep keypoint detector, and show that PACE3D# improves over the state of the art in vehicle pose estimation in the ApolloScape datasets, and its runtime is compatible with practical applications. We release our code at //github.com/MIT-SPARK/PACE.
Depth perception is considered an invaluable source of information in the context of 3D mapping and various robotics applications. However, point cloud maps acquired using consumer-level light detection and ranging sensors (lidars) still suffer from bias related to local surface properties such as measuring beam-to-surface incidence angle, distance, texture, reflectance, or illumination conditions. This fact has recently motivated researchers to exploit traditional filters, as well as the deep learning paradigm, in order to suppress the aforementioned depth sensors error while preserving geometric and map consistency details. Despite the effort, depth correction of lidar measurements is still an open challenge mainly due to the lack of clean 3D data that could be used as ground truth. In this paper, we introduce two novel point cloud map consistency losses, which facilitate self-supervised learning on real data of lidar depth correction models. Specifically, the models exploit multiple point cloud measurements of the same scene from different view-points in order to learn to reduce the bias based on the constructed map consistency signal. Complementary to the removal of the bias from the measurements, we demonstrate that the depth correction models help to reduce localization drift. Additionally, we release a data set that contains point cloud data captured in an indoor corridor environment with precise localization and ground truth mapping information.
Generative models, as an important family of statistical modeling, target learning the observed data distribution via generating new instances. Along with the rise of neural networks, deep generative models, such as variational autoencoders (VAEs) and generative adversarial network (GANs), have made tremendous progress in 2D image synthesis. Recently, researchers switch their attentions from the 2D space to the 3D space considering that 3D data better aligns with our physical world and hence enjoys great potential in practice. However, unlike a 2D image, which owns an efficient representation (i.e., pixel grid) by nature, representing 3D data could face far more challenges. Concretely, we would expect an ideal 3D representation to be capable enough to model shapes and appearances in details, and to be highly efficient so as to model high-resolution data with fast speed and low memory cost. However, existing 3D representations, such as point clouds, meshes, and recent neural fields, usually fail to meet the above requirements simultaneously. In this survey, we make a thorough review of the development of 3D generation, including 3D shape generation and 3D-aware image synthesis, from the perspectives of both algorithms and more importantly representations. We hope that our discussion could help the community track the evolution of this field and further spark some innovative ideas to advance this challenging task.
Generative models are now capable of producing highly realistic images that look nearly indistinguishable from the data on which they are trained. This raises the question: if we have good enough generative models, do we still need datasets? We investigate this question in the setting of learning general-purpose visual representations from a black-box generative model rather than directly from data. Given an off-the-shelf image generator without any access to its training data, we train representations from the samples output by this generator. We compare several representation learning methods that can be applied to this setting, using the latent space of the generator to generate multiple "views" of the same semantic content. We show that for contrastive methods, this multiview data can naturally be used to identify positive pairs (nearby in latent space) and negative pairs (far apart in latent space). We find that the resulting representations rival those learned directly from real data, but that good performance requires care in the sampling strategy applied and the training method. Generative models can be viewed as a compressed and organized copy of a dataset, and we envision a future where more and more "model zoos" proliferate while datasets become increasingly unwieldy, missing, or private. This paper suggests several techniques for dealing with visual representation learning in such a future. Code is released on our project page: //ali-design.github.io/GenRep/
Weakly-Supervised Object Detection (WSOD) and Localization (WSOL), i.e., detecting multiple and single instances with bounding boxes in an image using image-level labels, are long-standing and challenging tasks in the CV community. With the success of deep neural networks in object detection, both WSOD and WSOL have received unprecedented attention. Hundreds of WSOD and WSOL methods and numerous techniques have been proposed in the deep learning era. To this end, in this paper, we consider WSOL is a sub-task of WSOD and provide a comprehensive survey of the recent achievements of WSOD. Specifically, we firstly describe the formulation and setting of the WSOD, including the background, challenges, basic framework. Meanwhile, we summarize and analyze all advanced techniques and training tricks for improving detection performance. Then, we introduce the widely-used datasets and evaluation metrics of WSOD. Lastly, we discuss the future directions of WSOD. We believe that these summaries can help pave a way for future research on WSOD and WSOL.
Current deep learning research is dominated by benchmark evaluation. A method is regarded as favorable if it empirically performs well on the dedicated test set. This mentality is seamlessly reflected in the resurfacing area of continual learning, where consecutively arriving sets of benchmark data are investigated. The core challenge is framed as protecting previously acquired representations from being catastrophically forgotten due to the iterative parameter updates. However, comparison of individual methods is nevertheless treated in isolation from real world application and typically judged by monitoring accumulated test set performance. The closed world assumption remains predominant. It is assumed that during deployment a model is guaranteed to encounter data that stems from the same distribution as used for training. This poses a massive challenge as neural networks are well known to provide overconfident false predictions on unknown instances and break down in the face of corrupted data. In this work we argue that notable lessons from open set recognition, the identification of statistically deviating data outside of the observed dataset, and the adjacent field of active learning, where data is incrementally queried such that the expected performance gain is maximized, are frequently overlooked in the deep learning era. Based on these forgotten lessons, we propose a consolidated view to bridge continual learning, active learning and open set recognition in deep neural networks. Our results show that this not only benefits each individual paradigm, but highlights the natural synergies in a common framework. We empirically demonstrate improvements when alleviating catastrophic forgetting, querying data in active learning, selecting task orders, while exhibiting robust open world application where previously proposed methods fail.
We advocate the use of implicit fields for learning generative models of shapes and introduce an implicit field decoder for shape generation, aimed at improving the visual quality of the generated shapes. An implicit field assigns a value to each point in 3D space, so that a shape can be extracted as an iso-surface. Our implicit field decoder is trained to perform this assignment by means of a binary classifier. Specifically, it takes a point coordinate, along with a feature vector encoding a shape, and outputs a value which indicates whether the point is outside the shape or not. By replacing conventional decoders by our decoder for representation learning and generative modeling of shapes, we demonstrate superior results for tasks such as shape autoencoding, generation, interpolation, and single-view 3D reconstruction, particularly in terms of visual quality.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191