In this paper, we address the problem of plausible object placement for the challenging task of realistic image composition. We propose DiffPop, the first framework that utilizes plausibility-guided denoising diffusion probabilistic model to learn the scale and spatial relations among multiple objects and the corresponding scene image. First, we train an unguided diffusion model to directly learn the object placement parameters in a self-supervised manner. Then, we develop a human-in-the-loop pipeline which exploits human labeling on the diffusion-generated composite images to provide the weak supervision for training a structural plausibility classifier. The classifier is further used to guide the diffusion sampling process towards generating the plausible object placement. Experimental results verify the superiority of our method for producing plausible and diverse composite images on the new Cityscapes-OP dataset and the public OPA dataset, as well as demonstrate its potential in applications such as data augmentation and multi-object placement tasks. Our dataset and code will be released.
Recent advancements in generative models have enabled image editing very effective with impressive results. By extending this progress to 3D geometry models, we introduce iShapEditing, a novel framework for 3D shape editing which is applicable to both generated and real shapes. Users manipulate shapes by dragging handle points to corresponding targets, offering an intuitive and intelligent editing interface. Leveraging the Triplane Diffusion model and robust intermediate feature correspondence, our framework utilizes classifier guidance to adjust noise representations during sampling process, ensuring alignment with user expectations while preserving plausibility. For real shapes, we employ shape predictions at each time step alongside a DDPM-based inversion algorithm to derive their latent codes, facilitating seamless editing. iShapEditing provides effective and intelligent control over shapes without the need for additional model training or fine-tuning. Experimental examples demonstrate the effectiveness and superiority of our method in terms of editing accuracy and plausibility.
We introduce 𝒢-Style, a novel algorithm designed to transfer the style of an image onto a 3D scene represented using Gaussian Splatting. Gaussian Splatting is a powerful 3D representation for novel view synthesis, as—compared to other approaches based on Neural Radiance Fields—it provides fast scene renderings and user control over the scene. Recent pre-prints have demonstrated that the style of Gaussian Splatting scenes can be modified using an image exemplar. However, since the scene geometry remains fixed during the stylization process, current solutions fall short of producing satisfactory results. Our algorithm aims to address these limitations by following a three-step process: In a pre-processing step, we remove undesirable Gaussians with large projection areas or highly elongated shapes. Subsequently, we combine several losses carefully designed to preserve different scales of the style in the image, while maintaining as much as possible the integrity of the original scene content. During the stylization process and following the original design of Gaussian Splatting, we split Gaussians where additional detail is necessary within our scene by tracking the gradient of the stylized color. Our experiments demonstrate that 𝒢-Style generates high-quality stylizations within just a few minutes, outperforming existing methods both qualitatively and quantitatively.
We introduce LGSur-Net, an end-to-end deep learning architecture, engineered for the upsampling of sparse point clouds. LGSur-Net harnesses a trainable Gaussian local representation by positioning a series of Gaussian functions on an oriented plane, complemented by the optimization of individual covariance matrices. The integration of parametric factors allows for the encoding of the plane's rotational dynamics and Gaussian weightings into a linear transformation matrix. Then we extract the feature maps from the point cloud and its adjoining edges and learn the local Gaussian depictions to accurately model the shape's local geometry through an attention-based network. The Gaussian representation's inherent high-order continuity endows LGSur-Net with the natural ability to predict surface normals and support upsampling to any specified resolution. Comprehensive experiments validate that LGSur-Net efficiently learns from sparse data inputs, surpassing the performance of existing state-of-the-art upsampling methods. Our code is publicly available at https://github.com/Rangiant5b72/LGSur-Net.
Radiance field methods represent the state of the art in reconstructing complex scenes from multi-view photos. However, these reconstructions often suffer from one or both of the following limitations: First, they typically represent scenes in low dynamic range (LDR), which restricts their use to evenly lit environments and hinders immersive viewing experiences. Secondly, their reliance on a pinhole camera model, assuming all scene elements are in focus in the input images, presents practical challenges and complicates refocusing during novel-view synthesis. Addressing these limitations, we present a lightweight method based on 3D Gaussian Splatting that utilizes multi-view LDR images of a scene with varying exposure times, apertures, and focus distances as input to reconstruct a high-dynamic-range (HDR) radiance field. By incorporating analytical convolutions of Gaussians based on a thin-lens camera model as well as a tonemapping module, our reconstructions enable the rendering of HDR content with flexible refocusing capabilities. We demonstrate that our combined treatment of HDR and depth of field facilitates real-time cinematic rendering, outperforming the state of the art.
As a fundamental problem in computer vision, 3D point cloud registration (PCR) aims to seek the optimal transformation to align point cloud pairs. Meanwhile, the equivariance lies at the core of matching point clouds at arbitrary pose. In this paper, we propose GETr, a geometric equivariant transformer for PCR. By learning the point-wise orientations, we decouple the coordinate to the pose of the point clouds, which is the key to achieve equivariance in our framework. Then we utilize attention mechanism to learn the geometric features for superpoints matching, the proposed novel self-attention mechanism encodes the geometric information of point clouds. Finally, the coarse-to-fine manner is used to obtain high-quality correspondence for registration. Extensive experiments on both indoor and outdoor benchmarks demonstrate that our method outperforms various existing state-of-the-art methods.
We describe a method to convert 3D shapes into neural implicit form such that the shape is approximated in a guaranteed conservative manner. This means the input shape is strictly contained inside the neural implicit or, alternatively, vice versa. Such conservative approximations are of interest in a variety of applications, including collision detection, occlusion culling, or intersection testing. Our approach is the first to guarantee conservativeness in this context of neural implicits. We support input given as mesh, voxel set, or implicit function. Adaptive affine arithmetic is employed in the neural network fitting process, enabling the reasoning over infinite sets of points despite using a finite set of training data. Combined with an interior point style optimization approach this yields the desired guarantee.
This paper introduces a novel continual learning framework for synthesising novel views of multiple scenes, learning multiple 3D scenes incrementally, and updating the network parameters only with the training data of the upcoming new scene. We build on Neural Radiance Fields (NeRF), which uses multi-layer perceptron to model the density and radiance field of a scene as the implicit function. While NeRF and its extensions have shown a powerful capability of rendering photo-realistic novel views in a single 3D scene, managing these growing 3D NeRF assets efficiently is a new scientific problem. Very few works focus on the efficient representation or continuous learning capability of multiple scenes, which is crucial for the practical applications of NeRF. To achieve these goals, our key idea is to represent multiple scenes as the linear combination of a cross-scene weight matrix and a set of scene-specific weight matrices generated from a global parameter generator. Furthermore, we propose an uncertain surface knowledge distillation strategy to transfer the radiance field knowledge of previous scenes to the new model. Representing multiple 3D scenes with such weight matrices significantly reduces memory requirements. At the same time, the uncertain surface distillation strategy greatly overcomes the catastrophic forgetting problem and maintains the photo-realistic rendering quality of previous scenes. Experiments show that the proposed approach achieves state-of-the-art rendering quality of continual learning NeRF on NeRF-Synthetic, LLFF, and TanksAndTemples datasets while preserving extra low storage cost.
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：