Tianfu Wang

Huiguo He, Tianfu Wang, Huan Yang et al. · microsoft-research

We study the task of generating profitable Non-Fungible Token (NFT) images from user-input texts. Recent advances in diffusion models have shown great potential for image generation. However, existing works can fall short in generating visually-pleasing and highly-profitable NFT images, mainly due to the lack of 1) plentiful and fine-grained visual attribute prompts for an NFT image, and 2) effective optimization metrics for generating high-quality NFT images. To solve these challenges, we propose a Diffusion-based generation framework with Multiple Visual-Policies as rewards (i.e., Diffusion-MVP) for NFT images. The proposed framework consists of a large language model (LLM), a diffusion-based image generator, and a series of visual rewards by design. First, the LLM enhances a basic human input (such as "panda") by generating more comprehensive NFT-style prompts that include specific visual attributes, such as "panda with Ninja style and green background." Second, the diffusion-based image generator is fine-tuned using a large-scale NFT dataset to capture fine-grained image styles and accessory compositions of popular NFT elements. Third, we further propose to utilize multiple visual-policies as optimization goals, including visual rarity levels, visual aesthetic scores, and CLIP-based text-image relevances. This design ensures that our proposed Diffusion-MVP is capable of minting NFT images with high visual quality and market value. To facilitate this research, we have collected the largest publicly available NFT image dataset to date, consisting of 1.5 million high-quality images with corresponding texts and market values. Extensive experiments including objective evaluations and user studies demonstrate that our framework can generate NFT images showing more visually engaging elements and higher market value, compared with SOTA approaches.

Tianfu Wang, Menelaos Kanakis, Konrad Schindler et al.

Diffusion-based text-to-image models ignited immense attention from the vision community, artists, and content creators. Broad adoption of these models is due to significant improvement in the quality of generations and efficient conditioning on various modalities, not just text. However, lifting the rich generative priors of these 2D models into 3D is challenging. Recent works have proposed various pipelines powered by the entanglement of diffusion models and neural fields. We explore the power of pretrained 2D diffusion models and standard 3D neural radiance fields as independent, standalone tools and demonstrate their ability to work together in a non-learned fashion. Such modularity has the intrinsic advantage of eased partial upgrades, which became an important property in such a fast-paced domain. Our pipeline accepts any legacy renderable geometry, such as textured or untextured meshes, orchestrates the interaction between 2D generative refinement and 3D consistency enforcement tools, and outputs a painted input geometry in several formats. We conduct a large-scale study on a wide range of objects and categories from the ShapeNetSem dataset and demonstrate the advantages of our approach, both qualitatively and quantitatively. Project page: https://www.obukhov.ai/repainting_3d_assets

Jiazhang Wang, Tianfu Wang, Bingjie Xu et al.

Eye-tracking plays a crucial role in the development of virtual reality devices, neuroscience research, and psychology. Despite its significance in numerous applications, achieving an accurate, robust, and fast eye-tracking solution remains a considerable challenge for current state-of-the-art methods. While existing reflection-based techniques (e.g., "glint tracking") are considered to be very accurate, their performance is limited by their reliance on sparse 3D surface data acquired solely from the cornea surface. In this paper, we rethink the way how specular reflections can be used for eye tracking: We propose a novel method for accurate and fast evaluation of the gaze direction that exploits teachings from single-shot phase-measuring-deflectometry(PMD). In contrast to state-of-the-art reflection-based methods, our method acquires dense 3D surface information of both cornea and sclera within only one single camera frame (single-shot). For a typical measurement, we acquire $>3000 \times$ more surface reflection points ("glints") than conventional methods. We show the feasibility of our approach with experimentally evaluated gaze errors on a realistic model eye below only $0.12^\circ$. Moreover, we demonstrate quantitative measurements on real human eyes in vivo, reaching accuracy values between only $0.46^\circ$ and $0.97^\circ$.

Tianfu Wang, Jiazhang Wang, Oliver Cossairt et al.

Eye tracking is an important tool with a wide range of applications in Virtual, Augmented, and Mixed Reality (VR/AR/MR) technologies. State-of-the-art eye tracking methods are either reflection-based and track reflections of sparse point light sources, or image-based and exploit 2D features of the acquired eye image. In this work, we attempt to significantly improve reflection-based methods by utilizing pixel-dense deflectometric surface measurements in combination with optimization-based inverse rendering algorithms. Utilizing the known geometry of our deflectometric setup, we develop a differentiable rendering pipeline based on PyTorch3D that simulates a virtual eye under screen illumination. Eventually, we exploit the image-screen-correspondence information from the captured measurements to find the eye's rotation, translation, and shape parameters with our renderer via gradient descent. In general, our method does not require a specific pattern and can work with ordinary video frames of the main VR/AR/MR screen itself. We demonstrate real-world experiments with evaluated mean relative gaze errors below 0.45 degrees at a precision better than 0.11 degrees. Moreover, we show an improvement of 6X over a representative reflection-based state-of-the-art method in simulation.

Tianfu Wang, Guosheng Hu, Hongguang Wang

Estimating the pose of objects from images is a crucial task of 3D scene understanding, and recent approaches have shown promising results on very large benchmarks. However, these methods experience a significant performance drop when dealing with unseen objects. We believe that it results from the limited generalizability of image features. To address this problem, we have an in-depth analysis on the features of diffusion models, e.g. Stable Diffusion, which hold substantial potential for modeling unseen objects. Based on this analysis, we then innovatively introduce these diffusion features for object pose estimation. To achieve this, we propose three distinct architectures that can effectively capture and aggregate diffusion features of different granularity, greatly improving the generalizability of object pose estimation. Our approach outperforms the state-of-the-art methods by a considerable margin on three popular benchmark datasets, LM, O-LM, and T-LESS. In particular, our method achieves higher accuracy than the previous best arts on unseen objects: 97.9% vs. 93.5% on Unseen LM, 85.9% vs. 76.3% on Unseen O-LM, showing the strong generalizability of our method. Our code is released at https://github.com/Tianfu18/diff-feats-pose.

Tianfu Wang, Anton Obukhov, Konrad Schindler

Generative 3D Painting is among the top productivity boosters in high-resolution 3D asset management and recycling. Ever since text-to-image models became accessible for inference on consumer hardware, the performance of 3D Painting methods has consistently improved and is currently close to plateauing. At the core of most such models lies denoising diffusion in the latent space, an inherently time-consuming iterative process. Multiple techniques have been developed recently to accelerate generation and reduce sampling iterations by orders of magnitude. Designed for 2D generative imaging, these techniques do not come with recipes for lifting them into 3D. In this paper, we address this shortcoming by proposing a Latent Consistency Model (LCM) adaptation for the task at hand. We analyze the strengths and weaknesses of the proposed model and evaluate it quantitatively and qualitatively. Based on the Objaverse dataset samples study, our 3D painting method attains strong preference in all evaluations. Source code is available at https://github.com/kongdai123/consistency2.

Anjany Sekuboyina, Malek E. Husseini, Amirhossein Bayat et al.

Vertebral labelling and segmentation are two fundamental tasks in an automated spine processing pipeline. Reliable and accurate processing of spine images is expected to benefit clinical decision-support systems for diagnosis, surgery planning, and population-based analysis on spine and bone health. However, designing automated algorithms for spine processing is challenging predominantly due to considerable variations in anatomy and acquisition protocols and due to a severe shortage of publicly available data. Addressing these limitations, the Large Scale Vertebrae Segmentation Challenge (VerSe) was organised in conjunction with the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) in 2019 and 2020, with a call for algorithms towards labelling and segmentation of vertebrae. Two datasets containing a total of 374 multi-detector CT scans from 355 patients were prepared and 4505 vertebrae have individually been annotated at voxel-level by a human-machine hybrid algorithm (https://osf.io/nqjyw/, https://osf.io/t98fz/). A total of 25 algorithms were benchmarked on these datasets. In this work, we present the the results of this evaluation and further investigate the performance-variation at vertebra-level, scan-level, and at different fields-of-view. We also evaluate the generalisability of the approaches to an implicit domain shift in data by evaluating the top performing algorithms of one challenge iteration on data from the other iteration. The principal takeaway from VerSe: the performance of an algorithm in labelling and segmenting a spine scan hinges on its ability to correctly identify vertebrae in cases of rare anatomical variations. The content and code concerning VerSe can be accessed at: https://github.com/anjany/verse.

Yuru Jia, Lukas Hoyer, Shengyu Huang et al.

Large, pretrained latent diffusion models (LDMs) have demonstrated an extraordinary ability to generate creative content, specialize to user data through few-shot fine-tuning, and condition their output on other modalities, such as semantic maps. However, are they usable as large-scale data generators, e.g., to improve tasks in the perception stack, like semantic segmentation? We investigate this question in the context of autonomous driving, and answer it with a resounding "yes". We propose an efficient data generation pipeline termed DGInStyle. First, we examine the problem of specializing a pretrained LDM to semantically-controlled generation within a narrow domain. Second, we propose a Style Swap technique to endow the rich generative prior with the learned semantic control. Third, we design a Multi-resolution Latent Fusion technique to overcome the bias of LDMs towards dominant objects. Using DGInStyle, we generate a diverse dataset of street scenes, train a domain-agnostic semantic segmentation model on it, and evaluate the model on multiple popular autonomous driving datasets. Our approach consistently increases the performance of several domain generalization methods compared to the previous state-of-the-art methods. The source code and the generated dataset are available at https://dginstyle.github.io.

Bingxin Ke, Kevin Qu, Tianfu Wang et al.

The success of deep learning in computer vision over the past decade has hinged on large labeled datasets and strong pretrained models. In data-scarce settings, the quality of these pretrained models becomes crucial for effective transfer learning. Image classification and self-supervised learning have traditionally been the primary methods for pretraining CNNs and transformer-based architectures. Recently, the rise of text-to-image generative models, particularly those using denoising diffusion in a latent space, has introduced a new class of foundational models trained on massive, captioned image datasets. These models' ability to generate realistic images of unseen content suggests they possess a deep understanding of the visual world. In this work, we present Marigold, a family of conditional generative models and a fine-tuning protocol that extracts the knowledge from pretrained latent diffusion models like Stable Diffusion and adapts them for dense image analysis tasks, including monocular depth estimation, surface normals prediction, and intrinsic decomposition. Marigold requires minimal modification of the pre-trained latent diffusion model's architecture, trains with small synthetic datasets on a single GPU over a few days, and demonstrates state-of-the-art zero-shot generalization. Project page: https://marigoldcomputervision.github.io

Liwei Deng, Tianfu Wang, Yan Zhao et al.

Portfolio management is an important yet challenging task in AI for FinTech, which aims to allocate investors' budgets among different assets to balance the risk and return of an investment. In this study, we propose a general Multi-objectIve framework with controLLable rIsk for pOrtfolio maNagement (MILLION), which consists of two main phases, i.e., return-related maximization and risk control. Specifically, in the return-related maximization phase, we introduce two auxiliary objectives, i.e., return rate prediction, and return rate ranking, combined with portfolio optimization to remit the overfitting problem and improve the generalization of the trained model to future markets. Subsequently, in the risk control phase, we propose two methods, i.e., portfolio interpolation and portfolio improvement, to achieve fine-grained risk control and fast risk adaption to a user-specified risk level. For the portfolio interpolation method, we theoretically prove that the risk can be perfectly controlled if the to-be-set risk level is in a proper interval. In addition, we also show that the return rate of the adjusted portfolio after portfolio interpolation is no less than that of the min-variance optimization, as long as the model in the reward maximization phase is effective. Furthermore, the portfolio improvement method can achieve greater return rates while keeping the same risk level compared to portfolio interpolation. Extensive experiments are conducted on three real-world datasets. The results demonstrate the effectiveness and efficiency of the proposed framework.

Zejian Chen, Wei Zhuo, Tianfu Wang et al.

In this work, we propose a novel straightforward method for medical volume and sequence segmentation with limited annotations. To avert laborious annotating, the recent success of self-supervised learning(SSL) motivates the pre-training on unlabeled data. Despite its success, it is still challenging to adapt typical SSL methods to volume/sequence segmentation, due to their lack of mining on local semantic discrimination and rare exploitation on volume and sequence structures. Based on the continuity between slices/frames and the common spatial layout of organs across volumes/sequences, we introduced a novel bootstrap self-supervised representation learning method by leveraging the predictable possibility of neighboring slices. At the core of our method is a simple and straightforward dense self-supervision on the predictions of local representations and a strategy of predicting locals based on global context, which enables stable and reliable supervision for both global and local representation mining among volumes. Specifically, we first proposed an asymmetric network with an attention-guided predictor to enforce distance-specific prediction and supervision on slices within and across volumes/sequences. Secondly, we introduced a novel prototype-based foreground-background calibration module to enhance representation consistency. The two parts are trained jointly on labeled and unlabeled data. When evaluated on three benchmark datasets of medical volumes and sequences, our model outperforms existing methods with a large margin of 4.5\% DSC on ACDC, 1.7\% on Prostate, and 2.3\% on CAMUS. Intensive evaluations reveals the effectiveness and superiority of our method.

Xin Yang, Haoran Dou, Ruobing Huang et al.

Accurate standard plane (SP) localization is the fundamental step for prenatal ultrasound (US) diagnosis. Typically, dozens of US SPs are collected to determine the clinical diagnosis. 2D US has to perform scanning for each SP, which is time-consuming and operator-dependent. While 3D US containing multiple SPs in one shot has the inherent advantages of less user-dependency and more efficiency. Automatically locating SP in 3D US is very challenging due to the huge search space and large fetal posture variations. Our previous study proposed a deep reinforcement learning (RL) framework with an alignment module and active termination to localize SPs in 3D US automatically. However, termination of agent search in RL is important and affects the practical deployment. In this study, we enhance our previous RL framework with a newly designed adaptive dynamic termination to enable an early stop for the agent searching, saving at most 67% inference time, thus boosting the accuracy and efficiency of the RL framework at the same time. Besides, we validate the effectiveness and generalizability of our algorithm extensively on our in-house multi-organ datasets containing 433 fetal brain volumes, 519 fetal abdomen volumes, and 683 uterus volumes. Our approach achieves localization error of 2.52mm/10.26 degrees, 2.48mm/10.39 degrees, 2.02mm/10.48 degrees, 2.00mm/14.57 degrees, 2.61mm/9.71 degrees, 3.09mm/9.58 degrees, 1.49mm/7.54 degrees for the transcerebellar, transventricular, transthalamic planes in fetal brain, abdominal plane in fetal abdomen, and mid-sagittal, transverse and coronal planes in uterus, respectively. Experimental results show that our method is general and has the potential to improve the efficiency and standardization of US scanning.

Wei Wang, Xiang-Gen Xia, Chuanjiang He et al.

In this paper, we first present an arc based algorithm for fan-beam computed tomography (CT) reconstruction via applying Katsevich's helical CT formula to 2D fan-beam CT reconstruction. Then, we propose a new weighting function to deal with the redundant projection data. By extending the weighted arc based fan-beam algorithm to circle cone-beam geometry, we also obtain a new FDK-similar algorithm for circle cone-beam CT reconstruction. Experiments show that our methods can obtain higher PSNR and SSIM compared to the Parker-weighted conventional fan-beam algorithm and the FDK algorithm for super-short-scan trajectories.

Mohammad Hamghalam, Baiying Lei, Tianfu Wang

Structural magnetic resonance imaging (MRI) has been widely utilized for analysis and diagnosis of brain diseases. Automatic segmentation of brain tumors is a challenging task for computer-aided diagnosis due to low-tissue contrast in the tumor subregions. To overcome this, we devise a novel pixel-wise segmentation framework through a convolutional 3D to 2D MR patch conversion model to predict class labels of the central pixel in the input sliding patches. Precisely, we first extract 3D patches from each modality to calibrate slices through the squeeze and excitation (SE) block. Then, the output of the SE block is fed directly into subsequent bottleneck layers to reduce the number of channels. Finally, the calibrated 2D slices are concatenated to obtain multimodal features through a 2D convolutional neural network (CNN) for prediction of the central pixel. In our architecture, both local inter-slice and global intra-slice features are jointly exploited to predict class label of the central voxel in a given patch through the 2D CNN classifier. We implicitly apply all modalities through trainable parameters to assign weights to the contributions of each sequence for segmentation. Experimental results on the segmentation of brain tumors in multimodal MRI scans (BraTS'19) demonstrate that our proposed method can efficiently segment the tumor regions.

Wei Wang, Xiang-Gen Xia, Chuanjiang He et al.

In this paper, we first propose a variational model for the limited-angle computed tomography (CT) image reconstruction and then convert the model into an end-to-end deep network.We use the penalty method to solve the model and divide it into three iterative subproblems, where the first subproblem completes the sinograms by utilizing the prior information of sinograms in the frequency domain and the second refines the CT images by using the prior information of CT images in the spatial domain, and the last merges the outputs of the first two subproblems. In each iteration, we use the convolutional neural networks (CNNs) to approxiamte the solutions of the first two subproblems and, thus, obtain an end-to-end deep network for the limited-angle CT image reconstruction. Our network tackles both the sinograms and the CT images, and can simultaneously suppress the artifacts caused by the incomplete data and recover fine structural information in the CT images. Experimental results show that our method outperforms the existing algorithms for the limited-angle CT image reconstruction.

Mohammad Hamghalam, Baiying Lei, Tianfu Wang

Magnetic resonance imaging (MRI) provides varying tissue contrast images of internal organs based on a strong magnetic field. Despite the non-invasive advantage of MRI in frequent imaging, the low contrast MR images in the target area make tissue segmentation a challenging problem. This paper demonstrates the potential benefits of image-to-image translation techniques to generate synthetic high tissue contrast (HTC) images. Notably, we adopt a new cycle generative adversarial network (CycleGAN) with an attention mechanism to increase the contrast within underlying tissues. The attention block, as well as training on HTC images, guides our model to converge on certain tissues. To increase the resolution of HTC images, we employ multi-stage architecture to focus on one particular tissue as a foreground and filter out the irrelevant background in each stage. This multi-stage structure also alleviates the common artifacts of the synthetic images by decreasing the gap between source and target domains. We show the application of our method for synthesizing HTC images on brain MR scans, including glioma tumor. We also employ HTC MR images in both the end-to-end and two-stage segmentation structure to confirm the effectiveness of these images. The experiments over three competitive segmentation baselines on BraTS 2018 dataset indicate that incorporating the synthetic HTC images in the multi-modal segmentation framework improves the average Dice scores 0.8%, 0.6%, and 0.5% on the whole tumor, tumor core, and enhancing tumor, respectively, while eliminating one real MRI sequence from the segmentation procedure.

Wei Wang, Xiang-Gen Xia, Chuanjiang He et al.

A CT image can be well reconstructed when the sampling rate of the sinogram satisfies the Nyquist criteria and the sampled signal is noise-free. However, in practice, the sinogram is usually contaminated by noise, which degrades the quality of a reconstructed CT image. In this paper, we design a deep network for sinogram and CT image reconstruction. The network consists of two cascaded blocks that are linked by a filter backprojection (FBP) layer, where the former block is responsible for denoising and completing the sinograms while the latter is used to removing the noise and artifacts of the CT images. Experimental results show that the reconstructed CT images by our methods have the highest PSNR and SSIM in average compared to state of the art methods.

Xinzi He, Baiying Lei, Tianfu Wang

The accurate detection of lesion attributes is meaningful for both the computeraid diagnosis system and dermatologists decisions. However, unlike lesion segmentation and melenoma classification, there are few deep learning methods and literatures focusing on this task. Currently, the lesion attribute detection still remains challenging due to the extremely unbalanced class distribution and insufficient samples, as well as large intraclass and low interclass variations. To solve these problems, we propose a deep learning framework named superpixel attention network (SANet). Firstly, we segment input images into small regions and shuffle the obtained regions by the random shuttle mechanism (RSM). Secondly, we apply the SANet to capture discriminative features and reconstruct input images. Specifically, SANet contains two sub modules: superpixel average pooling and superpixel at tention module. We introduce a superpixel average pooling to reformulate the superpixel classification problem as a superpixel segmentation problem and a SAMis utilized to focus on discriminative superpixel regions and feature channels. Finally, we design a novel but effective loss, namely global balancing loss to address the serious data imbalance in ISIC 2018 Task 2 lesion attributes detection dataset. The proposed method achieves quite good performance on the ISIC 2018 Task 2 challenge.

Mohammad Hamghalam, Baiying Lei, Tianfu Wang

The magnetic resonance (MR) analysis of brain tumors is widely used for diagnosis and examination of tumor subregions. The overlapping area among the intensity distribution of healthy, enhancing, non-enhancing, and edema regions makes the automatic segmentation a challenging task. Here, we show that a convolutional neural network trained on high-contrast images can transform the intensity distribution of brain lesions in its internal subregions. Specifically, a generative adversarial network (GAN) is extended to synthesize high-contrast images. A comparison of these synthetic images and real images of brain tumor tissue in MR scans showed significant segmentation improvement and decreased the number of real channels for segmentation. The synthetic images are used as a substitute for real channels and can bypass real modalities in the multimodal brain tumor segmentation framework. Segmentation results on BraTS 2019 dataset demonstrate that our proposed approach can efficiently segment the tumor areas. In the end, we predict patient survival time based on volumetric features of the tumor subregions as well as the age of each case through several regression models.