Radiance field is an effective representation of 3D scenes, which has been widely adopted in novel-view synthesis and 3D reconstruction. It is still an open and challenging problem to evaluate the geometry, i.e., the density field, as the ground-truth is almost impossible to be obtained. One alternative indirect solution is to transform the density field into a point-cloud and compute its Chamfer Distance with the scanned ground-truth. However, many widely-used datasets have no point-cloud ground-truth since the scanning process along with the equipment is expensive and complicated. To this end, we propose a novel metric, named Inverse Mean Residual Color (IMRC), which can evaluate the geometry only with the observation images. Our key insight is that the better the geometry is, the lower-frequency the computed color field is. From this insight, given reconstructed density field and the observation images, we design a closed-form method to approximate the color field with low-frequency spherical harmonics and compute the inverse mean residual color. Then the higher the IMRC, the better the geometry. Qualitative and quantitative experimental results verify the effectiveness of our proposed IMRC metric. We also benchmark several state-of-the-art methods using IMRC to promote future related research.
相關內容
Neural Radiance Field (NeRF) research has attracted significant attention recently, with 3D modelling, virtual/augmented reality, and visual effects driving its application. While current NeRF implementations can produce high quality visual results, there is a conspicuous lack of reliable methods for evaluating them. Conventional image quality assessment methods and analytical metrics (e.g. PSNR, SSIM, LPIPS etc.) only provide approximate indicators of performance since they generalise the ability of the entire NeRF pipeline. Hence, in this paper, we propose a new test framework which isolates the neural rendering network from the NeRF pipeline and then performs a parametric evaluation by training and evaluating the NeRF on an explicit radiance field representation. We also introduce a configurable approach for generating representations specifically for evaluation purposes. This employs ray-casting to transform mesh models into explicit NeRF samples, as well as to "shade" these representations. Combining these two approaches, we demonstrate how different "tasks" (scenes with different visual effects or learning strategies) and types of networks (NeRFs and depth-wise implicit neural representations (INRs)) can be evaluated within this framework. Additionally, we propose a novel metric to measure task complexity of the framework which accounts for the visual parameters and the distribution of the spatial data. Our approach offers the potential to create a comparative objective evaluation framework for NeRF methods.
Neural radiance fields (NeRFs) are able to synthesize realistic novel views from multi-view images captured from distinct positions and perspectives. In NeRF's rendering pipeline, neural networks are used to represent a scene independently or transform queried learnable feature vector of a point to the expected color or density. With the aid of geometry guides either in occupancy grids or proposal networks, the number of neural network evaluations can be reduced from hundreds to dozens in the standard volume rendering framework. Instead of rendering yielded color after neural network evaluation, we propose to render the queried feature vectors of a ray first and then transform the rendered feature vector to the final pixel color by a neural network. This fundamental change to the standard volume rendering framework requires only one single neural network evaluation to render a pixel, which substantially lowers the high computational complexity of the rendering framework attributed to a large number of neural network evaluations. Consequently, we can use a comparably larger neural network to achieve a better rendering quality while maintaining the same training and rendering time costs. Our model achieves the state-of-the-art rendering quality on both synthetic and real-world datasets while requiring a training time of several minutes.
Despite the promising results of multi-view reconstruction, the recent neural rendering-based methods, such as implicit surface rendering (IDR) and volume rendering (NeuS), not only incur a heavy computational burden on training but also have the difficulties in disentangling the geometric and appearance. Although having achieved faster training speed than implicit representation and hash coding, the explicit voxel-based method obtains the inferior results on recovering surface. To address these challenges, we propose an effective mesh-based neural rendering approach, named FastMESH, which only samples at the intersection of ray and mesh. A coarse-to-fine scheme is introduced to efficiently extract the initial mesh by space carving. More importantly, we suggest a hexagonal mesh model to preserve surface regularity by constraining the second-order derivatives of vertices, where only low level of positional encoding is engaged for neural rendering. The experiments demonstrate that our approach achieves the state-of-the-art results on both reconstruction and novel view synthesis. Besides, we obtain 10-fold acceleration on training comparing to the implicit representation-based methods.
We present a method for generating high-quality watertight manifold meshes from multi-view input images. Existing volumetric rendering methods are robust in optimization but tend to generate noisy meshes with poor topology. Differentiable rasterization-based methods can generate high-quality meshes but are sensitive to initialization. Our method combines the benefits of both worlds; we take the geometry initialization obtained from neural volumetric fields, and further optimize the geometry as well as a compact neural texture representation with differentiable rasterizers. Through extensive experiments, we demonstrate that our method can generate accurate mesh reconstructions with faithful appearance that are comparable to previous volume rendering methods while being an order of magnitude faster in rendering. We also show that our generated mesh and neural texture reconstruction is compatible with existing graphics pipelines and enables downstream 3D applications such as simulation. Project page: //sarahweiii.github.io/neumanifold/
Neural Radiance Fields (NeRF) enable 3D scene reconstruction from 2D images and camera poses for Novel View Synthesis (NVS). Although NeRF can produce photorealistic results, it often suffers from overfitting to training views, leading to poor geometry reconstruction, especially in low-texture areas. This limitation restricts many important applications which require accurate geometry, such as extrapolated NVS, HD mapping and scene editing. To address this limitation, we propose a new method to improve NeRF's 3D structure using only RGB images and semantic maps. Our approach introduces a novel plane regularization based on Singular Value Decomposition (SVD), that does not rely on any geometric prior. In addition, we leverage the Structural Similarity Index Measure (SSIM) in our loss design to properly initialize the volumetric representation of NeRF. Quantitative and qualitative results show that our method outperforms popular regularization approaches in accurate geometry reconstruction for large-scale outdoor scenes and achieves SoTA rendering quality on the KITTI-360 NVS benchmark.
Image restoration is a long-standing low-level vision problem, e.g., deblurring and deraining. In the process of image restoration, it is necessary to consider not only the spatial details and contextual information of restoration to ensure the quality, but also the system complexity. Although many methods have been able to guarantee the quality of image restoration, the system complexity of the state-of-the-art (SOTA) methods is increasing as well. Motivated by this, we present a mixed hierarchy network that can balance these competing goals. Our main proposal is a mixed hierarchy architecture, that progressively recovers contextual information and spatial details from degraded images while we design intra-blocks to reduce system complexity. Specifically, our model first learns the contextual information using encoder-decoder architectures, and then combines them with high-resolution branches that preserve spatial detail. In order to reduce the system complexity of this architecture for convenient analysis and comparison, we replace or remove the nonlinear activation function with multiplication and use a simple network structure. In addition, we replace spatial convolution with global self-attention for the middle block of encoder-decoder. The resulting tightly interlinked hierarchy architecture, named as MHNet, delivers strong performance gains on several image restoration tasks, including image deraining, and deblurring.
Deep learning in computer vision has achieved great success with the price of large-scale labeled training data. However, exhaustive data annotation is impracticable for each task of all domains of interest, due to high labor costs and unguaranteed labeling accuracy. Besides, the uncontrollable data collection process produces non-IID training and test data, where undesired duplication may exist. All these nuisances may hinder the verification of typical theories and exposure to new findings. To circumvent them, an alternative is to generate synthetic data via 3D rendering with domain randomization. We in this work push forward along this line by doing profound and extensive research on bare supervised learning and downstream domain adaptation. Specifically, under the well-controlled, IID data setting enabled by 3D rendering, we systematically verify the typical, important learning insights, e.g., shortcut learning, and discover the new laws of various data regimes and network architectures in generalization. We further investigate the effect of image formation factors on generalization, e.g., object scale, material texture, illumination, camera viewpoint, and background in a 3D scene. Moreover, we use the simulation-to-reality adaptation as a downstream task for comparing the transferability between synthetic and real data when used for pre-training, which demonstrates that synthetic data pre-training is also promising to improve real test results. Lastly, to promote future research, we develop a new large-scale synthetic-to-real benchmark for image classification, termed S2RDA, which provides more significant challenges for transfer from simulation to reality. The code and datasets are available at //github.com/huitangtang/On_the_Utility_of_Synthetic_Data.
Image fusion plays a key role in a variety of multi-sensor-based vision systems, especially for enhancing visual quality and/or extracting aggregated features for perception. However, most existing methods just consider image fusion as an individual task, thus ignoring its underlying relationship with these downstream vision problems. Furthermore, designing proper fusion architectures often requires huge engineering labor. It also lacks mechanisms to improve the flexibility and generalization ability of current fusion approaches. To mitigate these issues, we establish a Task-guided, Implicit-searched and Meta-initialized (TIM) deep model to address the image fusion problem in a challenging real-world scenario. Specifically, we first propose a constrained strategy to incorporate information from downstream tasks to guide the unsupervised learning process of image fusion. Within this framework, we then design an implicit search scheme to automatically discover compact architectures for our fusion model with high efficiency. In addition, a pretext meta initialization technique is introduced to leverage divergence fusion data to support fast adaptation for different kinds of image fusion tasks. Qualitative and quantitative experimental results on different categories of image fusion problems and related downstream tasks (e.g., visual enhancement and semantic understanding) substantiate the flexibility and effectiveness of our TIM. The source code will be available at //github.com/LiuZhu-CV/TIMFusion.
Architectures that first convert point clouds to a grid representation and then apply convolutional neural networks achieve good performance for radar-based object detection. However, the transfer from irregular point cloud data to a dense grid structure is often associated with a loss of information, due to the discretization and aggregation of points. In this paper, we propose a novel architecture, multi-scale KPPillarsBEV, that aims to mitigate the negative effects of grid rendering. Specifically, we propose a novel grid rendering method, KPBEV, which leverages the descriptive power of kernel point convolutions to improve the encoding of local point cloud contexts during grid rendering. In addition, we propose a general multi-scale grid rendering formulation to incorporate multi-scale feature maps into convolutional backbones of detection networks with arbitrary grid rendering methods. We perform extensive experiments on the nuScenes dataset and evaluate the methods in terms of detection performance and computational complexity. The proposed multi-scale KPPillarsBEV architecture outperforms the baseline by 5.37% and the previous state of the art by 2.88% in Car AP4.0 (average precision for a matching threshold of 4 meters) on the nuScenes validation set. Moreover, the proposed single-scale KPBEV grid rendering improves the Car AP4.0 by 2.90% over the baseline while maintaining the same inference speed.
We present a monocular Simultaneous Localization and Mapping (SLAM) using high level object and plane landmarks, in addition to points. The resulting map is denser, more compact and meaningful compared to point only SLAM. We first propose a high order graphical model to jointly infer the 3D object and layout planes from single image considering occlusions and semantic constraints. The extracted cuboid object and layout planes are further optimized in a unified SLAM framework. Objects and planes can provide more semantic constraints such as Manhattan and object supporting relationships compared to points. Experiments on various public and collected datasets including ICL NUIM and TUM mono show that our algorithm can improve camera localization accuracy compared to state-of-the-art SLAM and also generate dense maps in many structured environments.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191