Liang Zhao

Xiaoxiao He, Chaowei Tan, Bo Liu et al.

Federated Learning has gained popularity among medical institutions since it enables collaborative training between clients (e.g., hospitals) without aggregating data. However, due to the high cost associated with creating annotations, especially for large 3D image datasets, clinical institutions do not have enough supervised data for training locally. Thus, the performance of the collaborative model is subpar under limited supervision. On the other hand, large institutions have the resources to compile data repositories with high-resolution images and labels. Therefore, individual clients can utilize the knowledge acquired in the public data repositories to mitigate the shortage of private annotated images. In this paper, we propose a federated few-shot learning method with dual knowledge distillation. This method allows joint training with limited annotations across clients without jeopardizing privacy. The supervised learning of the proposed method extracts features from limited labeled data in each client, while the unsupervised data is used to distill both feature and response-based knowledge from a national data repository to further improve the accuracy of the collaborative model and reduce the communication cost. Extensive evaluations are conducted on 3D magnetic resonance knee images from a private clinical dataset. Our proposed method shows superior performance and less training time than other semi-supervised federated learning methods. Codes and additional visualization results are available at https://github.com/hexiaoxiao-cs/fedml-knee.

Haoran Wei, Lingyu Kong, Jinyue Chen et al.

Modern Large Vision-Language Models (LVLMs) enjoy the same vision vocabulary -- CLIP, which can cover most common vision tasks. However, for some special vision task that needs dense and fine-grained vision perception, e.g., document-level OCR or chart understanding, especially in non-English scenarios, the CLIP-style vocabulary may encounter low efficiency in tokenizing the vision knowledge and even suffer out-of-vocabulary problem. Accordingly, we propose Vary, an efficient and effective method to scale up the vision vocabulary of LVLMs. The procedures of Vary are naturally divided into two folds: the generation and integration of a new vision vocabulary. In the first phase, we devise a vocabulary network along with a tiny decoder-only transformer to produce the desired vocabulary via autoregression. In the next, we scale up the vanilla vision vocabulary by merging the new one with the original one (CLIP), enabling the LVLMs can quickly garner new features. Compared to the popular BLIP-2, MiniGPT4, and LLaVA, Vary can maintain its vanilla capabilities while enjoying more excellent fine-grained perception and understanding ability. Specifically, Vary is competent in new document parsing features (OCR or markdown conversion) while achieving 78.2% ANLS in DocVQA and 36.2% in MMVet. Our code will be publicly available on the homepage.

Guangji Bai, Qilong Zhao, Xiaoyang Jiang et al.

Continual Learning is a burgeoning domain in next-generation AI, focusing on training neural networks over a sequence of tasks akin to human learning. While CL provides an edge over traditional supervised learning, its central challenge remains to counteract catastrophic forgetting and ensure the retention of prior tasks during subsequent learning. Amongst various strategies to tackle this, replay based methods have emerged as preeminent, echoing biological memory mechanisms. However, these methods are memory intensive, often preserving entire data samples, an approach inconsistent with humans selective memory retention of salient experiences. While some recent works have explored the storage of only significant portions of data in episodic memory, the inherent nature of partial data necessitates innovative retrieval mechanisms. Current solutions, like inpainting, approximate full data reconstruction from partial cues, a method that diverges from genuine human memory processes. Addressing these nuances, this paper presents the Saliency Guided Hidden Associative Replay for Continual Learning. This novel framework synergizes associative memory with replay-based strategies. SHARC primarily archives salient data segments via sparse memory encoding. Importantly, by harnessing associative memory paradigms, it introduces a content focused memory retrieval mechanism, promising swift and near-perfect recall, bringing CL a step closer to authentic human memory processes. Extensive experimental results demonstrate the effectiveness of our proposed method for various continual learning tasks.

Ke Zhu, Zheng Ge, Liang Zhao et al.

This paper makes the first attempt towards unsupervised preference alignment in Vision-Language Models (VLMs). We generate chosen and rejected responses with regard to the original and augmented image pairs, and conduct preference alignment with direct preference optimization. It is based on a core idea: properly designed augmentation to the image input will induce VLM to generate false but hard negative responses, which helps the model to learn from and produce more robust and powerful answers. The whole pipeline no longer hinges on supervision from GPT-4 or human involvement during alignment, and is highly efficient with few lines of code. With only 8k randomly sampled unsupervised data, it achieves 90\% relative score to GPT-4 on complex reasoning in LLaVA-Bench, and improves LLaVA-7B/13B by 6.7\%/5.6\% score on complex multi-modal benchmark MM-Vet. Visualizations shows its improved ability to align with user-intentions. A series of ablations are firmly conducted to reveal the latent mechanism of the approach, which also indicates its potential towards further scaling. Code are available in https://github.com/Kevinz-code/SeVa.

Xiangru Li, Yifei Zhang, Liang Zhao

The Segment Anything Model (SAM) is a powerful foundation model that introduced revolutionary advancements in natural image segmentation. However, its performance remains sub-optimal when delineating the intricate structure of biomedical images, where multiple organs and tissues intertwine in a single image. In this study, we introduce a novel fine-tuning framework that leverages SAM's ability to bundle and process multiple prompts per image and seeks to improve SAM's performance in medical images. We first curated a medical image dataset that consists of CT scans of lesions in various organs, each with two annotations for organs and lesions respectively. Then, we fine-tuned SAM's mask decoder within our framework by batching both bounding boxes generated from ground truth masks as reference. The batched prompt strategy we introduced not only addresses the inherent complexity and ambiguity often found in medical images but also substantially enhances performance metrics when applied onto a wide range of segmentation tasks.

Yiyang Ma, Xingchao Liu, Xiaokang Chen et al.

We present JanusFlow, a powerful framework that unifies image understanding and generation in a single model. JanusFlow introduces a minimalist architecture that integrates autoregressive language models with rectified flow, a state-of-the-art method in generative modeling. Our key finding demonstrates that rectified flow can be straightforwardly trained within the large language model framework, eliminating the need for complex architectural modifications. To further improve the performance of our unified model, we adopt two key strategies: (i) decoupling the understanding and generation encoders, and (ii) aligning their representations during unified training. Extensive experiments show that JanusFlow achieves comparable or superior performance to specialized models in their respective domains, while significantly outperforming existing unified approaches across standard benchmarks. This work represents a step toward more efficient and versatile vision-language models.

Bo Pan, Zhen Xiong, Guanchen Wu et al.

Graph representation learning has garnered significant attention due to its broad applications in various domains, such as recommendation systems and social network analysis. Despite advancements in graph learning methods, challenges still remain in explainability when graphs are associated with semantic features. In this paper, we present GraphNarrator, the first method designed to generate natural language explanations for Graph Neural Networks. GraphNarrator employs a generative language model that maps input-output pairs to explanations reflecting the model's decision-making process. To address the lack of ground truth explanations to train the model, we propose first generating pseudo-labels that capture the model's decisions from saliency-based explanations, then using Expert Iteration to iteratively train the pseudo-label generator based on training objectives on explanation quality. The high-quality pseudo-labels are finally utilized to train an end-to-end explanation generator model. Extensive experiments are conducted to demonstrate the effectiveness of GraphNarrator in producing faithful, concise, and human-preferred natural language explanations.

Mengdan Zhu, Senhao Cheng, Guangji Bai et al.

Text-to-image generation increasingly demands access to domain-specific, fine-grained, and rapidly evolving knowledge that pretrained models cannot fully capture, necessitating the integration of retrieval methods. Existing Retrieval-Augmented Generation (RAG) methods attempt to address this by retrieving globally relevant images, but they fail when no single image contains all desired elements from a complex user query. We propose Cross-modal RAG, a novel framework that decomposes both queries and images into sub-dimensional components, enabling subquery-aware retrieval and generation. Our method introduces a hybrid retrieval strategy - combining a sub-dimensional sparse retriever with a dense retriever - to identify a Pareto-optimal set of images, each contributing complementary aspects of the query. During generation, a multimodal large language model is guided to selectively condition on relevant visual features aligned to specific subqueries, ensuring subquery-aware image synthesis. Extensive experiments on MS-COCO, Flickr30K, WikiArt, CUB, and ImageNet-LT demonstrate that Cross-modal RAG significantly outperforms existing baselines in the retrieval and further contributes to generation quality, while maintaining high efficiency.

Yiran Zhou, Yingyu Wang, Shoudong Huang et al.

Multiview point cloud registration is a fundamental task for constructing globally consistent 3D models. Existing approaches typically rely on feature extraction and data association across multiple point clouds; however, these processes are challenging to obtain global optimal solution in complex environments. In this paper, we introduce a novel correspondence-free multiview point cloud registration method. Specifically, we represent the global map as a depth map and leverage raw depth information to formulate a non-linear least squares optimisation that jointly estimates poses of point clouds and the global map. Unlike traditional feature-based bundle adjustment methods, which rely on explicit feature extraction and data association, our method bypasses these challenges by associating multi-frame point clouds with a global depth map through their corresponding poses. This data association is implicitly incorporated and dynamically refined during the optimisation process. Extensive evaluations on real-world datasets demonstrate that our method outperforms state-of-the-art approaches in accuracy, particularly in challenging environments where feature extraction and data association are difficult.

Yang Song, Zhuqing Zhang, Jun Wu et al.

With the recent advance of deep learning based object recognition and estimation, it is possible to consider object level SLAM where the pose of each object is estimated in the SLAM process. In this paper, based on a novel Lie group structure, a right invariant extended Kalman filter (RI-EKF) for object based SLAM is proposed. The observability analysis shows that the proposed algorithm automatically maintains the correct unobservable subspace, while standard EKF (Std-EKF) based SLAM algorithm does not. This results in a better consistency for the proposed algorithm comparing to Std-EKF. Finally, simulations and real world experiments validate not only the consistency and accuracy of the proposed algorithm, but also the practicability of the proposed RI-EKF for object based SLAM problem. The MATLAB code of the algorithm is made publicly available.

Zhehua Mao, Liang Zhao, Shoudong Huang et al.

This paper presents a novel algorithm named Direct Simultaneous Registration (DSR) that registers a collection of 3D images in a simultaneous fashion without specifying any reference image, feature extraction and matching, or information loss or reuse. The algorithm optimizes the global poses of local image frames by maximizing the similarity between a predefined panoramic image and local images. Although we formulate the problem as a Direct Bundle Adjustment (DBA) that jointly optimizes the poses of local frames and the intensities of the panoramic image, by investigating the independence of pose estimation from the panoramic image in the solving process, DSR is proposed to solve the poses only and proved to be able to obtain the same optimal poses as DBA. The proposed method is particularly suitable for the scenarios where distinct features are not available, such as Transesophageal Echocardiography (TEE) images. DSR is evaluated by comparing it with four widely used methods via simulated and in-vivo 3D TEE images. It is shown that the proposed method outperforms these four methods in terms of accuracy and requires much fewer computational resources than the state-of-the-art accumulated pairwise estimates (APE).

Jingwei Song, Jun Wang, Liang Zhao et al.

In minimal invasive surgery, it is important to rebuild and visualize the latest deformed shape of soft-tissue surfaces to mitigate tissue damages. This paper proposes an innovative Simultaneous Localization and Mapping (SLAM) algorithm for deformable dense reconstruction of surfaces using a sequence of images from a stereoscope. We introduce a warping field based on the Embedded Deformation (ED) nodes with 3D shapes recovered from consecutive pairs of stereo images. The warping field is estimated by deforming the last updated model to the current live model. Our SLAM system can: (1) Incrementally build a live model by progressively fusing new observations with vivid accurate texture. (2) Estimate the deformed shape of unobserved region with the principle As-Rigid-As-Possible. (3) Show the consecutive shape of models. (4) Estimate the current relative pose between the soft-tissue and the scope. In-vivo experiments with publicly available datasets demonstrate that the 3D models can be incrementally built for different soft-tissues with different deformations from sequences of stereo images obtained by laparoscopes. Results show the potential clinical application of our SLAM system for providing surgeon useful shape and texture information in minimal invasive surgery.

Felix H. Kong, Jiaheng Zhao, Liang Zhao et al.

In this paper, we show that for a minimal pose-graph problem, even in the ideal case of perfect measurements and spherical covariance, using the so-called "wrap function" when comparing angles results in multiple suboptimal local minima. We numerically estimate regions of attraction to these local minima for some numerical examples, and give evidence to show that they are of nonzero measure. In contrast, under the same assumptions, we show that the \textit{chordal distance} representation of angle error has a unique minimum up to periodicity. For chordal cost, we also search for initial conditions that fail to converge to the global minimum, and find that this occurs with far fewer points than with geodesic cost.

Jingwei Song, Liang Zhao, Shoudong Huang et al.

In this paper, we study the back-end of simultaneous localization and mapping (SLAM) problem in deforming environment, where robot localizes itself and tracks multiple non-rigid soft surface using its onboard sensor measurements. An elaborate analysis is conducted on conventional deformation modelling method, Embedded Deformation (ED) graph. We demonstrate and prove that the ED graph widely used in such scenarios is unobservable and leads to multiple solutions unless suitable priors are provided. Example as well as theoretical prove are provided to show the ambiguity of ED graph and camera pose. In modelling non-rigid scenario with ED graph, motion priors of the deforming environment is essential to separate robot pose and deforming environment. The conclusion can be extrapolated to any free form deformation formulation. In solving the observability, this research proposes a preliminary deformable SLAM approach to estimate robot pose in complex environments that exhibits regular motion. A strategy that approximates deformed shape using a linear combination of several previous shapes is proposed to avoid the ambiguity in robot movement and rigid and non-rigid motions of the environment. Fisher information matrix rank analysis with a base case is discussed to prove the effectiveness. Moreover, the proposed algorithm is validated extensively on Monte Carlo simulations and real experiments. It is demonstrated that the new algorithm significantly outperforms conventional rigid SLAM and ED based SLAM especially in scenarios where there is large deformation.

Jingwei Song, Fang Bai, Liang Zhao et al.

Embedded deformation nodes based formulation has been widely applied in deformable geometry and graphical problems. Though being promising in stereo (or RGBD) sensor based SLAM applications, it remains challenging to keep constant speed in deformation nodes parameter estimation when model grows larger. In practice, the processing time grows rapidly in accordance with the expansion of maps. In this paper, we propose an approach to decouple nodes of deformation graph in large scale dense deformable SLAM and keep the estimation time to be constant. We observe that only partial deformable nodes in the graph are connected to visible points. Based on this fact, sparsity of original Hessian matrix is utilized to split parameter estimation in two independent steps. With this new technique, we achieve faster parameter estimation with amortized computation complexity reduced from O(n^2) to closing O(1). As a result, the computation cost barely increases as the map keeps growing. Based on our strategy, computational bottleneck in large scale embedded deformation graph based applications will be greatly mitigated. The effectiveness is validated by experiments, featuring large scale deformation scenarios.

Liang Zhao, Shoudong Huang, Gamini Dissanayake

The main contribution of this paper is a new submap joining based approach for solving large-scale Simultaneous Localization and Mapping (SLAM) problems. Each local submap is independently built using the local information through solving a small-scale SLAM; the joining of submaps mainly involves solving linear least squares and performing nonlinear coordinate transformations. Through approximating the local submap information as the state estimate and its corresponding information matrix, judiciously selecting the submap coordinate frames, and approximating the joining of a large number of submaps by joining only two maps at a time, either sequentially or in a more efficient Divide and Conquer manner, the nonlinear optimization process involved in most of the existing submap joining approaches is avoided. Thus the proposed submap joining algorithm does not require initial guess or iterations since linear least squares problems have closed-form solutions. The proposed Linear SLAM technique is applicable to feature-based SLAM, pose graph SLAM and D-SLAM, in both two and three dimensions, and does not require any assumption on the character of the covariance matrices. Simulations and experiments are performed to evaluate the proposed Linear SLAM algorithm. Results using publicly available datasets in 2D and 3D show that Linear SLAM produces results that are very close to the best solutions that can be obtained using full nonlinear optimization algorithm started from an accurate initial guess. The C/C++ and MATLAB source codes of Linear SLAM are available on OpenSLAM.

Liyang Liu, Teng Zhang, Yi Liu et al.

Bundle adjustment (BA) with parallax angle based feature parameterization has been shown to have superior performance over BA using inverse depth or XYZ feature forms. In this paper, we propose an improved version of the parallax BA algorithm (PMBA) by extending it to the manifold domain along with observation-ray based objective function. With this modification, the problem formulation faithfully mimics the projective nature in a camera's image formation, BA is able to achieve better convergence, accuracy and robustness. This is particularly useful in handling diverse outdoor environments and collinear motion modes. Capitalizing on these properties, we further propose a pose-graph simplification to PMBA, with significant dimensionality reduction. This pose-graph model is convex in nature, easy to solve and its solution can serve as a good initial guess to the original BA problem which is intrinsically non-convex. We provide theoretical proof that our global initialization strategy can guarantee a near-optimal solution. Using a series of experiments involving diverse environmental conditions and motions, we demonstrate PMBA's superior convergence performance in comparison to other BA methods. We also show that, without incremental initialization or via third-party information, our global initialization process helps to bootstrap the full BA successfully in various scenarios, sequential or out-of-order, including some datasets from the "Bundle Adjustment in the Large" database.

Jingwei Song, Jun Wang, Liang Zhao et al.

Real-time simultaneously localization and dense mapping is very helpful for providing Virtual Reality and Augmented Reality for surgeons or even surgical robots. In this paper, we propose MIS-SLAM: a complete real-time large scale dense deformable SLAM system with stereoscope in Minimal Invasive Surgery based on heterogeneous computing by making full use of CPU and GPU. Idled CPU is used to perform ORB- SLAM for providing robust global pose. Strategies are taken to integrate modules from CPU and GPU. We solved the key problem raised in previous work, that is, fast movement of scope and blurry images make the scope tracking fail. Benefiting from improved localization, MIS-SLAM can achieve large scale scope localizing and dense mapping in real-time. It transforms and deforms current model and incrementally fuses new observation while keeping vivid texture. In-vivo experiments conducted on publicly available datasets presented in the form of videos demonstrate the feasibility and practicality of MIS-SLAM for potential clinical purpose.