Yafei Song

Feng Wang, Sinan Tan, Xinghang Li et al. · tsinghua

Synthesizing high-fidelity videos from real-world multi-view input is challenging because of the complexities of real-world environments and highly dynamic motions. Previous works based on neural radiance fields have demonstrated high-quality reconstructions of dynamic scenes. However, training such models on real-world scenes is time-consuming, usually taking days or weeks. In this paper, we present a novel method named MixVoxels to better represent the dynamic scenes with fast training speed and competitive rendering qualities. The proposed MixVoxels represents the 4D dynamic scenes as a mixture of static and dynamic voxels and processes them with different networks. In this way, the computation of the required modalities for static voxels can be processed by a lightweight model, which essentially reduces the amount of computation, especially for many daily dynamic scenes dominated by the static background. To separate the two kinds of voxels, we propose a novel variation field to estimate the temporal variance of each voxel. For the dynamic voxels, we design an inner-product time query method to efficiently query multiple time steps, which is essential to recover the high-dynamic motions. As a result, with 15 minutes of training for dynamic scenes with inputs of 300-frame videos, MixVoxels achieves better PSNR than previous methods. Codes and trained models are available at https://github.com/fengres/mixvoxels

Qihang Fang, Yafei Song, Keqiang Li et al.

A 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 obtain. 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, the lower-frequency the computed color field. From this insight, given a reconstructed density field and 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. Our code is available at https://github.com/qihangGH/IMRC.

Feng Wang, Zilong Chen, Guokang Wang et al.

In this paper, we propose the Masked Space-Time Hash encoding (MSTH), a novel method for efficiently reconstructing dynamic 3D scenes from multi-view or monocular videos. Based on the observation that dynamic scenes often contain substantial static areas that result in redundancy in storage and computations, MSTH represents a dynamic scene as a weighted combination of a 3D hash encoding and a 4D hash encoding. The weights for the two components are represented by a learnable mask which is guided by an uncertainty-based objective to reflect the spatial and temporal importance of each 3D position. With this design, our method can reduce the hash collision rate by avoiding redundant queries and modifications on static areas, making it feasible to represent a large number of space-time voxels by hash tables with small size.Besides, without the requirements to fit the large numbers of temporally redundant features independently, our method is easier to optimize and converge rapidly with only twenty minutes of training for a 300-frame dynamic scene.As a result, MSTH obtains consistently better results than previous methods with only 20 minutes of training time and 130 MB of memory storage. Code is available at https://github.com/masked-spacetime-hashing/msth

Yongxiang Liu, Jie Zhou, Yafei Song et al.

Synthetic Aperture Radar (SAR) imagery generation is essential for deepening the study of scattering mechanisms, establishing trustworthy electromagnetic scene models, and fundamentally alleviating the data scarcity bottleneck that constrains development in this field. However, existing methods find it difficult to simultaneously ensure high fidelity in both global geospatial semantics and microscopic scattering mechanisms, resulting in severe challenges for global generation. To address this, we propose \textbf{HuiYanEarth-SAR}, the first foundational SAR imagery generation model based on AlphaEarth and integrated scattering mechanisms. By injecting geospatial priors to control macroscopic structures and utilizing implicit scattering characteristic modeling to ensure the authenticity of microscopic textures, we achieve the capability of generating high-fidelity SAR images for global locations solely based on geographic coordinates. This study not only constructs an efficient SAR scene simulator but also establishes a bridge connecting geography, scatter mechanism, and artificial intelligence from a methodological standpoint. It advances SAR research by expanding the paradigm from perception and understanding to simulation and creation, providing key technical support for constructing a high-confidence digital twin of the Earth.

Qihang Fang, Yafei Song, Keqiang Li et al.

Neural radiance field (NeRF) enables the synthesis of cutting-edge realistic novel view images of a 3D scene. It includes density and color fields to model the shape and radiance of a scene, respectively. Supervised by the photometric loss in an end-to-end training manner, NeRF inherently suffers from the shape-radiance ambiguity problem, i.e., it can perfectly fit training views but does not guarantee decoupling the two fields correctly. To deal with this issue, existing works have incorporated prior knowledge to provide an independent supervision signal for the density field, including total variation loss, sparsity loss, distortion loss, etc. These losses are based on general assumptions about the density field, e.g., it should be smooth, sparse, or compact, which are not adaptive to a specific scene. In this paper, we propose a more adaptive method to reduce the shape-radiance ambiguity. The key is a rendering method that is only based on the density field. Specifically, we first estimate the color field based on the density field and posed images in a closed form. Then NeRF's rendering process can proceed. We address the problems in estimating the color field, including occlusion and non-uniformly distributed views. Afterward, it is applied to regularize NeRF's density field. As our regularization is guided by photometric loss, it is more adaptive compared to existing ones. Experimental results show that our method improves the density field of NeRF both qualitatively and quantitatively. Our code is available at https://github.com/qihangGH/Closed-form-color-field.

Yafei Song, Peng Zhang, Bang Zhang

Synthesizing synchronized and natural co-speech gesture videos remains a formidable challenge. Recent approaches have leveraged motion graphs to harness the potential of existing video data. To retrieve an appropriate trajectory from the graph, previous methods either utilize the distance between features extracted from the input audio and those associated with the motions in the graph or embed both the input audio and motion into a shared feature space. However, these techniques may not be optimal due to the many-to-many mapping nature between audio and gestures, which cannot be adequately addressed by one-to-one mapping. To alleviate this limitation, we propose a novel framework that initially employs a diffusion model to generate gesture motions. The diffusion model implicitly learns the joint distribution of audio and motion, enabling the generation of contextually appropriate gestures from input audio sequences. Furthermore, our method extracts both low-level and high-level features from the input audio to enrich the training process of the diffusion model. Subsequently, a meticulously designed motion-based retrieval algorithm is applied to identify the most suitable path within the graph by assessing both global and local similarities in motion. Given that not all nodes in the retrieved path are sequentially continuous, the final step involves seamlessly stitching together these segments to produce a coherent video output. Experimental results substantiate the efficacy of our proposed method, demonstrating a significant improvement over prior approaches in terms of synchronization accuracy and naturalness of generated gestures.

Yongxiang Liu, Weijie Li, Li Liu et al.

The absence of publicly available, large-scale, high-quality datasets for Synthetic Aperture Radar Automatic Target Recognition (SAR ATR) has significantly hindered the application of rapidly advancing deep learning techniques, which hold huge potential to unlock new capabilities in this field. This is primarily because collecting large volumes of diverse target samples from SAR images is prohibitively expensive, largely due to privacy concerns, the characteristics of microwave radar imagery perception, and the need for specialized expertise in data annotation. Throughout the history of SAR ATR research, there have been only a number of small datasets, mainly including targets like ships, airplanes, buildings, etc. There is only one vehicle dataset MSTAR collected in the 1990s, which has been a valuable source for SAR ATR. To fill this gap, this paper introduces a large-scale, new dataset named ATRNet-STAR with 40 different vehicle categories collected under various realistic imaging conditions and scenes. It marks a substantial advancement in dataset scale and diversity, comprising over 190,000 well-annotated samples, 10 times larger than its predecessor, the famous MSTAR. Building such a large dataset is a challenging task, and the data collection scheme will be detailed. Secondly, we illustrate the value of ATRNet-STAR via extensively evaluating the performance of 15 representative methods with 7 different experimental settings on challenging classification and detection benchmarks derived from the dataset. Finally, based on our extensive experiments, we identify valuable insights for SAR ATR and discuss potential future research directions in this field. We hope that the scale, diversity, and benchmark of ATRNet-STAR can significantly facilitate the advancement of SAR ATR.

Jianqiang Ren, Chao He, Lin Liu et al.

There is a growing demand for customized and expressive 3D characters with the emergence of AI agents and Metaverse, but creating 3D characters using traditional computer graphics tools is a complex and time-consuming task. To address these challenges, we propose a user-friendly framework named Make-A-Character (Mach) to create lifelike 3D avatars from text descriptions. The framework leverages the power of large language and vision models for textual intention understanding and intermediate image generation, followed by a series of human-oriented visual perception and 3D generation modules. Our system offers an intuitive approach for users to craft controllable, realistic, fully-realized 3D characters that meet their expectations within 2 minutes, while also enabling easy integration with existing CG pipeline for dynamic expressiveness. For more information, please visit the project page at https://human3daigc.github.io/MACH/.

Xin Gao, Li Hu, Siqi Hu et al.

Current state-of-the-art (SOTA) methods for audio-driven character animation demonstrate promising performance for scenarios primarily involving speech and singing. However, they often fall short in more complex film and television productions, which demand sophisticated elements such as nuanced character interactions, realistic body movements, and dynamic camera work. To address this long-standing challenge of achieving film-level character animation, we propose an audio-driven model, which we refere to as Wan-S2V, built upon Wan. Our model achieves significantly enhanced expressiveness and fidelity in cinematic contexts compared to existing approaches. We conducted extensive experiments, benchmarking our method against cutting-edge models such as Hunyuan-Avatar and Omnihuman. The experimental results consistently demonstrate that our approach significantly outperforms these existing solutions. Additionally, we explore the versatility of our method through its applications in long-form video generation and precise video lip-sync editing.

Gang Cheng, Xin Gao, Li Hu et al.

We introduce Wan-Animate, a unified framework for character animation and replacement. Given a character image and a reference video, Wan-Animate can animate the character by precisely replicating the expressions and movements of the character in the video to generate high-fidelity character videos. Alternatively, it can integrate the animated character into the reference video to replace the original character, replicating the scene's lighting and color tone to achieve seamless environmental integration. Wan-Animate is built upon the Wan model. To adapt it for character animation tasks, we employ a modified input paradigm to differentiate between reference conditions and regions for generation. This design unifies multiple tasks into a common symbolic representation. We use spatially-aligned skeleton signals to replicate body motion and implicit facial features extracted from source images to reenact expressions, enabling the generation of character videos with high controllability and expressiveness. Furthermore, to enhance environmental integration during character replacement, we develop an auxiliary Relighting LoRA. This module preserves the character's appearance consistency while applying the appropriate environmental lighting and color tone. Experimental results demonstrate that Wan-Animate achieves state-of-the-art performance. We are committed to open-sourcing the model weights and its source code.

Yafei Song, Ling Cai, Jia Li et al.

Researchers have attempted utilizing deep neural network (DNN) to learn novel local features from images inspired by its recent successes on a variety of vision tasks. However, existing DNN-based algorithms have not achieved such remarkable progress that could be partly attributed to insufficient utilization of the interactive characters between local feature detector and descriptor. To alleviate these difficulties, we emphasize two desired properties, i.e., repeatability and reliability, to simultaneously summarize the inherent and interactive characters of local feature detector and descriptor. Guided by these properties, a self-supervised framework, namely self-evolving keypoint detection and description (SEKD), is proposed to learn an advanced local feature model from unlabeled natural images. Additionally, to have performance guarantees, novel training strategies have also been dedicatedly designed to minimize the gap between the learned feature and its properties. We benchmark the proposed method on homography estimation, relative pose estimation, and structure-from-motion tasks. Extensive experimental results demonstrate that the proposed method outperforms popular hand-crafted and DNN-based methods by remarkable margins. Ablation studies also verify the effectiveness of each critical training strategy. We will release our code along with the trained model publicly.

Hongwei Xie, Shuo Zhang, Huanghao Ding et al.

The inherent heavy computation of deep neural networks prevents their widespread applications. A widely used method for accelerating model inference is quantization, by replacing the input operands of a network using fixed-point values. Then the majority of computation costs focus on the integer matrix multiplication accumulation. In fact, high-bit accumulator leads to partially wasted computation and low-bit one typically suffers from numerical overflow. To address this problem, we propose an overflow aware quantization method by designing trainable adaptive fixed-point representation, to optimize the number of bits for each input tensor while prohibiting numeric overflow during the computation. With the proposed method, we are able to fully utilize the computing power to minimize the quantization loss and obtain optimized inference performance. To verify the effectiveness of our method, we conduct image classification, object detection, and semantic segmentation tasks on ImageNet, Pascal VOC, and COCO datasets, respectively. Experimental results demonstrate that the proposed method can achieve comparable performance with state-of-the-art quantization methods while accelerating the inference process by about 2 times.

Yafei Song, Jia Li, Xiaogang Wang et al.

Single image dehazing, which aims to recover the clear image solely from an input hazy or foggy image, is a challenging ill-posed problem. Analysing existing approaches, the common key step is to estimate the haze density of each pixel. To this end, various approaches often heuristically designed haze-relevant features. Several recent works also automatically learn the features via directly exploiting Convolutional Neural Networks (CNN). However, it may be insufficient to fully capture the intrinsic attributes of hazy images. To obtain effective features for single image dehazing, this paper presents a novel Ranking Convolutional Neural Network (Ranking-CNN). In Ranking-CNN, a novel ranking layer is proposed to extend the structure of CNN so that the statistical and structural attributes of hazy images can be simultaneously captured. By training Ranking-CNN in a well-designed manner, powerful haze-relevant features can be automatically learned from massive hazy image patches. Based on these features, haze can be effectively removed by using a haze density prediction model trained through the random forest regression. Experimental results show that our approach outperforms several previous dehazing approaches on synthetic and real-world benchmark images. Comprehensive analyses are also conducted to interpret the proposed Ranking-CNN from both the theoretical and experimental aspects.

Yafei Song, Di Zhu, Jia Li et al.

In recent years, camera-based localization has been widely used for robotic applications, and most proposed algorithms rely on local features extracted from recorded images. For better performance, the features used for open-loop localization are required to be short-term globally static, and the ones used for re-localization or loop closure detection need to be long-term static. Therefore, the motion attribute of a local feature point could be exploited to improve localization performance, e.g., the feature points extracted from moving persons or vehicles can be excluded from these systems due to their unsteadiness. In this paper, we design a fully convolutional network (FCN), named MD-Net, to perform motion attribute estimation and feature description simultaneously. MD-Net has a shared backbone network to extract features from the input image and two network branches to complete each sub-task. With MD-Net, we can obtain the motion attribute while avoiding increasing much more computation. Experimental results demonstrate that the proposed method can learn distinct local feature descriptor along with motion attribute only using an FCN, by outperforming competing methods by a wide margin. We also show that the proposed algorithm can be integrated into a vision-based localization algorithm to improve estimation accuracy significantly.

Kui Fu, Peipei Shi, Yafei Song et al.

Large convolutional neural network models have recently demonstrated impressive performance on video attention prediction. Conventionally, these models are with intensive computation and large memory. To address these issues, we design an extremely light-weight network with ultrafast speed, named UVA-Net. The network is constructed based on depth-wise convolutions and takes low-resolution images as input. However, this straight-forward acceleration method will decrease performance dramatically. To this end, we propose a coupled knowledge distillation strategy to augment and train the network effectively. With this strategy, the model can further automatically discover and emphasize implicit useful cues contained in the data. Both spatial and temporal knowledge learned by the high-resolution complex teacher networks also can be distilled and transferred into the proposed low-resolution light-weight spatiotemporal network. Experimental results show that the performance of our model is comparable to 11 state-of-the-art models in video attention prediction, while it costs only 0.68 MB memory footprint, runs about 10,106 FPS on GPU and 404 FPS on CPU, which is 206 times faster than previous models.

Yafei Song, Yonghong Tian, Gang Wang et al.

It is a significant problem to predict the 2D LiDAR map at next moment for robotics navigation and path-planning. To tackle this problem, we resort to the motion flow between adjacent maps, as motion flow is a powerful tool to process and analyze the dynamic data, which is named optical flow in video processing. However, unlike video, which contains abundant visual features in each frame, a 2D LiDAR map lacks distinctive local features. To alleviate this challenge, we propose to estimate the motion flow based on deep neural networks inspired by its powerful representation learning ability in estimating the optical flow of the video. To this end, we design a recurrent neural network based on gated recurrent unit, which is named LiDAR-FlowNet. As a recurrent neural network can encode the temporal dynamic information, our LiDAR-FlowNet can estimate motion flow between the current map and the unknown next map only from the current frame and previous frames. A self-supervised strategy is further designed to train the LiDAR-FlowNet model effectively, while no training data need to be manually annotated. With the estimated motion flow, it is straightforward to predict the 2D LiDAR map at the next moment. Experimental results verify the effectiveness of our LiDAR-FlowNet as well as the proposed training strategy. The results of the predicted LiDAR map also show the advantages of our motion flow based method.

Jia Li, Junjie Wu, Anlin Zheng et al.

Segmenting primary objects in a video is an important yet challenging problem in computer vision, as it exhibits various levels of foreground/background ambiguities. To reduce such ambiguities, we propose a novel formulation via exploiting foreground and background context as well as their complementary constraint. Under this formulation, a unified objective function is further defined to encode each cue. For implementation, we design a Complementary Segmentation Network (CSNet) with two separate branches, which can simultaneously encode the foreground and background information along with joint spatial constraints. The CSNet is trained on massive images with manually annotated salient objects in an end-to-end manner. By applying CSNet on each video frame, the spatial foreground and background maps can be initialized. To enforce temporal consistency effectively and efficiently, we divide each frame into superpixels and construct neighborhood reversible flow that reflects the most reliable temporal correspondences between superpixels in far-away frames. With such flow, the initialized foregroundness and backgroundness can be propagated along the temporal dimension so that primary video objects gradually pop-out and distractors are well suppressed. Extensive experimental results on three video datasets show that the proposed approach achieves impressive performance in comparisons with 18 state-of-the-art models.

Jia Li, Yafei Song, Jianfeng Zhu et al.

Many advances of deep learning techniques originate from the efforts of addressing the image classification task on large-scale datasets. However, the construction of such clean datasets is costly and time-consuming since the Internet is overwhelmed by noisy images with inadequate and inaccurate tags. In this paper, we propose a Ubiquitous Reweighting Network (URNet) that learns an image classification model from large-scale noisy data. By observing the web data, we find that there are five key challenges, i.e., imbalanced class sizes, high intra-classes diversity and inter-class similarity, imprecise instances, insufficient representative instances, and ambiguous class labels. To alleviate these challenges, we assume that every training instance has the potential to contribute positively by alleviating the data bias and noise via reweighting the influence of each instance according to different class sizes, large instance clusters, its confidence, small instance bags and the labels. In this manner, the influence of bias and noise in the web data can be gradually alleviated, leading to the steadily improving performance of URNet. Experimental results in the WebVision 2018 challenge with 16 million noisy training images from 5000 classes show that our approach outperforms state-of-the-art models and ranks the first place in the image classification task.