Qiyu Sun

Chaoqiang Zhao, Matteo Poggi, Fabio Tosi et al.

This paper tackles the challenges of self-supervised monocular depth estimation in indoor scenes caused by large rotation between frames and low texture. We ease the learning process by obtaining coarse camera poses from monocular sequences through multi-view geometry to deal with the former. However, we found that limited by the scale ambiguity across different scenes in the training dataset, a naïve introduction of geometric coarse poses cannot play a positive role in performance improvement, which is counter-intuitive. To address this problem, we propose to refine those poses during training through rotation and translation/scale optimization. To soften the effect of the low texture, we combine the global reasoning of vision transformers with an overfitting-aware, iterative self-distillation mechanism, providing more accurate depth guidance coming from the network itself. Experiments on NYUv2, ScanNet, 7scenes, and KITTI datasets support the effectiveness of each component in our framework, which sets a new state-of-the-art for indoor self-supervised monocular depth estimation, as well as outstanding generalization ability. Code and models are available at https://github.com/zxcqlf/GasMono

Ruihao Xia, Chaoqiang Zhao, Meng Zheng et al.

Most nighttime semantic segmentation studies are based on domain adaptation approaches and image input. However, limited by the low dynamic range of conventional cameras, images fail to capture structural details and boundary information in low-light conditions. Event cameras, as a new form of vision sensors, are complementary to conventional cameras with their high dynamic range. To this end, we propose a novel unsupervised Cross-Modality Domain Adaptation (CMDA) framework to leverage multi-modality (Images and Events) information for nighttime semantic segmentation, with only labels on daytime images. In CMDA, we design the Image Motion-Extractor to extract motion information and the Image Content-Extractor to extract content information from images, in order to bridge the gap between different modalities (Images to Events) and domains (Day to Night). Besides, we introduce the first image-event nighttime semantic segmentation dataset. Extensive experiments on both the public image dataset and the proposed image-event dataset demonstrate the effectiveness of our proposed approach. We open-source our code, models, and dataset at https://github.com/XiaRho/CMDA.

Wenqi Ren, Yang Tang, Qiyu Sun et al.

Visual semantic segmentation aims at separating a visual sample into diverse blocks with specific semantic attributes and identifying the category for each block, and it plays a crucial role in environmental perception. Conventional learning-based visual semantic segmentation approaches count heavily on large-scale training data with dense annotations and consistently fail to estimate accurate semantic labels for unseen categories. This obstruction spurs a craze for studying visual semantic segmentation with the assistance of few/zero-shot learning. The emergence and rapid progress of few/zero-shot visual semantic segmentation make it possible to learn unseen-category from a few labeled or zero-labeled samples, which advances the extension to practical applications. Therefore, this paper focuses on the recently published few/zero-shot visual semantic segmentation methods varying from 2D to 3D space and explores the commonalities and discrepancies of technical settlements under different segmentation circumstances. Specifically, the preliminaries on few/zero-shot visual semantic segmentation, including the problem definitions, typical datasets, and technical remedies, are briefly reviewed and discussed. Moreover, three typical instantiations are involved to uncover the interactions of few/zero-shot learning with visual semantic segmentation, including image semantic segmentation, video object segmentation, and 3D segmentation. Finally, the future challenges of few/zero-shot visual semantic segmentation are discussed.

Qiyu Sun, Gary G. Yen, Yang Tang et al.

Monocular depth estimation is one of the fundamental tasks in environmental perception and has achieved tremendous progress in virtue of deep learning. However, the performance of trained models tends to degrade or deteriorate when employed on other new datasets due to the gap between different datasets. Though some methods utilize domain adaptation technologies to jointly train different domains and narrow the gap between them, the trained models cannot generalize to new domains that are not involved in training. To boost the transferability of depth estimation models, we propose an adversarial depth estimation task and train the model in the pipeline of meta-learning. Our proposed adversarial task mitigates the issue of meta-overfitting, since the network is trained in an adversarial manner and aims to extract domain invariant representations. In addition, we propose a constraint to impose upon cross-task depth consistency to compel the depth estimation to be identical in different adversarial tasks, which improves the performance of our method and smoothens the training process. Experiments demonstrate that our method adapts well to new datasets after few training steps during the test procedure.

Wenqi Ren, Qiyu Sun, Chaoqiang Zhao et al.

Learning-based image dehazing methods are essential to assist autonomous systems in enhancing reliability. Due to the domain gap between synthetic and real domains, the internal information learned from synthesized images is usually sub-optimal in real domains, leading to severe performance drop of dehaizing models. Driven by the ability on exploring internal information from a few unseen-domain samples, meta-learning is commonly adopted to address this issue via test-time training, which is hyperparameter-sensitive and time-consuming. In contrast, we present a domain generalization framework based on meta-learning to dig out representative and discriminative internal properties of real hazy domains without test-time training. To obtain representative domain-specific information, we attach two entities termed adaptation network and distance-aware aggregator to our dehazing network. The adaptation network assists in distilling domain-relevant information from a few hazy samples and caching it into a collection of features. The distance-aware aggregator strives to summarize the generated features and filter out misleading information for more representative internal properties. To enhance the discrimination of distilled internal information, we present a novel loss function called domain-relevant contrastive regularization, which encourages the internal features generated from the same domain more similar and that from diverse domains more distinct. The generated representative and discriminative features are regarded as some external variables of our dehazing network to regress a particular and powerful function for a given domain. The extensive experiments on real hazy datasets, such as RTTS and URHI, validate that our proposed method has superior generalization ability than the state-of-the-art competitors.

Qiyu Sun, Huilin Chen, Meng Zheng et al.

Domain generalized semantic segmentation (DGSS) is a critical yet challenging task, where the model is trained only on source data without access to any target data. Despite the proposal of numerous DGSS strategies, the generalization capability remains limited in CNN architectures. Though some Transformer-based segmentation models show promising performance, they primarily focus on capturing intra-sample attentive relationships, disregarding inter-sample correlations which can potentially benefit DGSS. To this end, we enhance the attention modules in Transformer networks for improving DGSS by incorporating information from other independent samples in the same batch, enriching contextual information, and diversifying the training data for each attention block. Specifically, we propose two alternative intra-batch attention mechanisms, namely mean-based intra-batch attention (MIBA) and element-wise intra-batch attention (EIBA), to capture correlations between different samples, enhancing feature representation and generalization capabilities. Building upon intra-batch attention, we introduce IBAFormer, which integrates self-attention modules with the proposed intra-batch attention for DGSS. Extensive experiments demonstrate that IBAFormer achieves SOTA performance in DGSS, and ablation studies further confirm the effectiveness of each introduced component.

Kexuan Zhang, Qiyu Sun, Chaoqiang Zhao et al.

Deep learning has revolutionized the field of artificial intelligence. Based on the statistical correlations uncovered by deep learning-based methods, computer vision has contributed to tremendous growth in areas like autonomous driving and robotics. Despite being the basis of deep learning, such correlation is not stable and is susceptible to uncontrolled factors. In the absence of the guidance of prior knowledge, statistical correlations can easily turn into spurious correlations and cause confounders. As a result, researchers are now trying to enhance deep learning methods with causal theory. Causal theory models the intrinsic causal structure unaffected by data bias and is effective in avoiding spurious correlations. This paper aims to comprehensively review the existing causal methods in typical vision and vision-language tasks such as semantic segmentation, object detection, and image captioning. The advantages of causality and the approaches for building causal paradigms will be summarized. Future roadmaps are also proposed, including facilitating the development of causal theory and its application in other complex scenes and systems.

Tianyu Wu, Yang Tang, Qiyu Sun et al.

In recent years, artificial intelligence has played an important role on accelerating the whole process of drug discovery. Various of molecular representation schemes of different modals (e.g. textual sequence or graph) are developed. By digitally encoding them, different chemical information can be learned through corresponding network structures. Molecular graphs and Simplified Molecular Input Line Entry System (SMILES) are popular means for molecular representation learning in current. Previous works have done attempts by combining both of them to solve the problem of specific information loss in single-modal representation on various tasks. To further fusing such multi-modal imformation, the correspondence between learned chemical feature from different representation should be considered. To realize this, we propose a novel framework of molecular joint representation learning via Multi-Modal information of SMILES and molecular Graphs, called MMSG. We improve the self-attention mechanism by introducing bond level graph representation as attention bias in Transformer to reinforce feature correspondence between multi-modal information. We further propose a Bidirectional Message Communication Graph Neural Network (BMC GNN) to strengthen the information flow aggregated from graphs for further combination. Numerous experiments on public property prediction datasets have demonstrated the effectiveness of our model.

Qiyu Sun, Pavlo Melnyk, Michael Felsberg et al.

Domain generalized semantic segmentation (DGSS) is an essential but highly challenging task, in which the model is trained only on source data and any target data is not available. Existing DGSS methods primarily standardize the feature distribution or utilize extra domain data for augmentation. However, the former sacrifices valuable information and the latter introduces domain biases. Therefore, generating diverse-style source data without auxiliary data emerges as an attractive strategy. In light of this, we propose GAN-based feature augmentation (GBFA) that hallucinates stylized feature maps while preserving their semantic contents with a feature generator. The impressive generative capability of GANs enables GBFA to perform inter-channel and trainable feature synthesis in an end-to-end framework. To enable learning GBFA, we introduce random image color augmentation (RICA), which adds a diverse range of variations to source images during training. These augmented images are then passed through a feature extractor to obtain features tailored for GBFA training. Both GBFA and RICA operate exclusively within the source domain, eliminating the need for auxiliary datasets. We conduct extensive experiments, and the generalization results from the synthetic GTAV and SYNTHIA to the real Cityscapes, BDDS, and Mapillary datasets show that our method achieves state-of-the-art performance in DGSS.

Huilin Chen, Qiyu Sun, Fangfei Li et al.

Computer vision tasks are crucial for aerospace missions as they help spacecraft to understand and interpret the space environment, such as estimating position and orientation, reconstructing 3D models, and recognizing objects, which have been extensively studied to successfully carry out the missions. However, traditional methods like Kalman Filtering, Structure from Motion, and Multi-View Stereo are not robust enough to handle harsh conditions, leading to unreliable results. In recent years, deep learning (DL)-based perception technologies have shown great potential and outperformed traditional methods, especially in terms of their robustness to changing environments. To further advance DL-based aerospace perception, various frameworks, datasets, and strategies have been proposed, indicating significant potential for future applications. In this survey, we aim to explore the promising techniques used in perception tasks and emphasize the importance of DL-based aerospace perception. We begin by providing an overview of aerospace perception, including classical space programs developed in recent years, commonly used sensors, and traditional perception methods. Subsequently, we delve into three fundamental perception tasks in aerospace missions: pose estimation, 3D reconstruction, and recognition, as they are basic and crucial for subsequent decision-making and control. Finally, we discuss the limitations and possibilities in current research and provide an outlook on future developments, including the challenges of working with limited datasets, the need for improved algorithms, and the potential benefits of multi-source information fusion.

Zixiang Wang, Mengjia Gong, Qiyu Sun et al.

With the rapid advancement of artificial intelligence, multi-agent systems (MASs) are evolving from classical paradigms toward architectures built upon large foundation models (LFMs). This survey provides a systematic review and comparative analysis of classical MASs (CMASs) and LFM-based MASs (LMASs). First, within a closed-loop coordination framework, CMASs are reviewed across four fundamental dimensions: perception, communication, decision-making, and control. Beyond this framework, LMASs integrate LFMs to lift collaboration from low-level state exchanges to semantic-level reasoning, enabling more flexible coordination and improved adaptability across diverse scenarios. Then, a comparative analysis is conducted to contrast CMASs and LMASs across architecture, operating mechanism, adaptability, and application. Finally, future perspectives on MASs are presented, summarizing open challenges and potential research opportunities.

Arash Amini, Qiyu Sun, Nader Motee

We consider a class of stochastic dynamical networks whose governing dynamics can be modeled using a coupling function. It is shown that the dynamics of such networks can generate geometrically ergodic trajectories under some reasonable assumptions. We show that a general class of coupling functions can be learned using only one sample trajectory from the network. This is practically plausible as in numerous applications it is desired to run an experiment only once but for a longer period of time, rather than repeating the same experiment multiple times from different initial conditions. Building upon ideas from the concentration inequalities for geometrically ergodic Markov chains, we formulate several results about the convergence of the empirical estimator to the true coupling function. Our theoretical findings are supported by extensive simulation results.

Fanshuang Kong, Richong Zhang, Qiyu Sun et al.

Mixup generates augmented samples by linearly interpolating inputs and labels with a controllable ratio. However, since it operates in the latent embedding level, the resulting samples are not human-interpretable. In contrast, LLM-based augmentation methods produce sentences via prompts at the token level, yielding readable outputs but offering limited control over the generation process. Inspired by recent advances in LLM inversion, which reconstructs natural language from embeddings and helps bridge the gap between latent embedding space and discrete token space, we propose inversedMixup, a unified framework that combines the controllability of Mixup with the interpretability of LLM-based generation. Specifically, inversedMixup adopts a three-stage training procedure to align the output embedding space of a task-specific model with the input embedding space of an LLM. Upon successful alignment, inversedMixup can reconstruct mixed embeddings with a controllable mixing ratio into human-interpretable augmented sentences, thereby improving the augmentation performance. Additionally, inversedMixup provides the first empirical evidence of the manifold intrusion phenomenon in text Mixup and introduces a simple yet effective strategy to mitigate it. Extensive experiments demonstrate the effectiveness and generalizability of our approach in both few-shot and fully supervised scenarios.

Ziyang Miao, Qiyu Sun, Jingyuan Wang et al.

Large language models (LLMs) have shown impressive performance on general-purpose tasks, yet adapting them to specific domains remains challenging due to the scarcity of high-quality domain data. Existing data synthesis tools often struggle to extract reliable fine-tuning data from heterogeneous documents effectively. To address this limitation, we propose Easy Dataset, a unified framework for synthesizing fine-tuning data from unstructured documents via an intuitive graphical user interface (GUI). Specifically, Easy Dataset allows users to easily configure text extraction models and chunking strategies to transform raw documents into coherent text chunks. It then leverages a persona-driven prompting approach to generate diverse question-answer pairs using public-available LLMs. Throughout the pipeline, a human-in-the-loop visual interface facilitates the review and refinement of intermediate outputs to ensure data quality. Experiments on a financial question-answering task show that fine-tuning LLMs on the synthesized dataset significantly improves domain-specific performance while preserving general knowledge. The source code and installable package are available at https://github.com/ConardLi/easy-dataset and have garnered over 9,000 GitHub stars.

Johan Edstedt, Qiyu Sun, Georg Bökman et al.

Feature matching is an important computer vision task that involves estimating correspondences between two images of a 3D scene, and dense methods estimate all such correspondences. The aim is to learn a robust model, i.e., a model able to match under challenging real-world changes. In this work, we propose such a model, leveraging frozen pretrained features from the foundation model DINOv2. Although these features are significantly more robust than local features trained from scratch, they are inherently coarse. We therefore combine them with specialized ConvNet fine features, creating a precisely localizable feature pyramid. To further improve robustness, we propose a tailored transformer match decoder that predicts anchor probabilities, which enables it to express multimodality. Finally, we propose an improved loss formulation through regression-by-classification with subsequent robust regression. We conduct a comprehensive set of experiments that show that our method, RoMa, achieves significant gains, setting a new state-of-the-art. In particular, we achieve a 36% improvement on the extremely challenging WxBS benchmark. Code is provided at https://github.com/Parskatt/RoMa

Seok-Young Chung, Qiyu Sun

Graph convolutional neural network (GCNN) operates on graph domain and it has achieved a superior performance to accomplish a wide range of tasks. In this paper, we introduce a Barron space of functions on a compact domain of graph signals. We prove that the proposed Barron space is a reproducing kernel Banach space, it can be decomposed into the union of a family of reproducing kernel Hilbert spaces with neuron kernels, and it could be dense in the space of continuous functions on the domain. Approximation property is one of the main principles to design neural networks. In this paper, we show that outputs of GCNNs are contained in the Barron space and functions in the Barron space can be well approximated by outputs of some GCNNs in the integrated square and uniform measurements. We also estimate the Rademacher complexity of functions with bounded Barron norm and conclude that functions in the Barron space could be learnt from their random samples efficiently.

Lu Liu, Yang Tang, Kexuan Zhang et al.

Due to the challenge posed by multi-source and heterogeneous data collected from diverse environments, causal relationships among features can exhibit variations influenced by different time spans, regions, or strategies. This diversity makes a single causal model inadequate for accurately representing complex causal relationships in all observational data, a crucial consideration in causal learning. To address this challenge, the nonlinear Causal Kernel Clustering method is introduced for heterogeneous subgroup causal learning, highlighting variations in causal relationships across diverse subgroups. The main component for clustering heterogeneous subgroups lies in the construction of the $u$-centered sample mapping function with the property of unbiased estimation, which assesses the differences in potential nonlinear causal relationships in various samples and supported by causal identifiability theory. Experimental results indicate that the method performs well in identifying heterogeneous subgroups and enhancing causal learning, leading to a reduction in prediction error.

Chaoqiang Zhao, Yang Tang, Qiyu Sun

With the development of computational intelligence algorithms, unsupervised monocular depth and pose estimation framework, which is driven by warped photometric consistency, has shown great performance in the daytime scenario. While in some challenging environments, like night and rainy night, the essential photometric consistency hypothesis is untenable because of the complex lighting and reflection, so that the above unsupervised framework cannot be directly applied to these complex scenarios. In this paper, we investigate the problem of unsupervised monocular depth estimation in highly complex scenarios and address this challenging problem by adopting an image transfer-based domain adaptation framework. We adapt the depth model trained on day-time scenarios to be applicable to night-time scenarios, and constraints on both feature space and output space promote the framework to learn the key features for depth decoding. Meanwhile, we further tackle the effects of unstable image transfer quality on domain adaptation, and an image adaptation approach is proposed to evaluate the quality of transferred images and re-weight the corresponding losses, so as to improve the performance of the adapted depth model. Extensive experiments show the effectiveness of the proposed unsupervised framework in estimating the dense depth map from highly complex images.

Chongzhen Zhang, Yang Tang, Chaoqiang Zhao et al.

Semantic segmentation and depth completion are two challenging tasks in scene understanding, and they are widely used in robotics and autonomous driving. Although several works are proposed to jointly train these two tasks using some small modifications, like changing the last layer, the result of one task is not utilized to improve the performance of the other one despite that there are some similarities between these two tasks. In this paper, we propose multi-task generative adversarial networks (Multi-task GANs), which are not only competent in semantic segmentation and depth completion, but also improve the accuracy of depth completion through generated semantic images. In addition, we improve the details of generated semantic images based on CycleGAN by introducing multi-scale spatial pooling blocks and the structural similarity reconstruction loss. Furthermore, considering the inner consistency between semantic and geometric structures, we develop a semantic-guided smoothness loss to improve depth completion results. Extensive experiments on Cityscapes dataset and KITTI depth completion benchmark show that the Multi-task GANs are capable of achieving competitive performance for both semantic segmentation and depth completion tasks.

Chaoqiang Zhao, Gary G. Yen, Qiyu Sun et al.

Previous work has shown that adversarial learning can be used for unsupervised monocular depth and visual odometry (VO) estimation, in which the adversarial loss and the geometric image reconstruction loss are utilized as the mainly supervisory signals to train the whole unsupervised framework. However, the performance of the adversarial framework and image reconstruction is usually limited by occlusions and the visual field changes between frames. This paper proposes a masked generative adversarial network (GAN) for unsupervised monocular depth and ego-motion estimation.The MaskNet and Boolean mask scheme are designed in this framework to eliminate the effects of occlusions and impacts of visual field changes on the reconstruction loss and adversarial loss, respectively. Furthermore, we also consider the scale consistency of our pose network by utilizing a new scale-consistency loss, and therefore, our pose network is capable of providing the full camera trajectory over a long monocular sequence. Extensive experiments on the KITTI dataset show that each component proposed in this paper contributes to the performance, and both our depth and trajectory predictions achieve competitive performance on the KITTI and Make3D datasets.

Chongzhen Zhang, Jianrui Wang, Gary G. Yen et al.

With widespread applications of artificial intelligence (AI), the capabilities of the perception, understanding, decision-making and control for autonomous systems have improved significantly in the past years. When autonomous systems consider the performance of accuracy and transferability, several AI methods, like adversarial learning, reinforcement learning (RL) and meta-learning, show their powerful performance. Here, we review the learning-based approaches in autonomous systems from the perspectives of accuracy and transferability. Accuracy means that a well-trained model shows good results during the testing phase, in which the testing set shares a same task or a data distribution with the training set. Transferability means that when a well-trained model is transferred to other testing domains, the accuracy is still good. Firstly, we introduce some basic concepts of transfer learning and then present some preliminaries of adversarial learning, RL and meta-learning. Secondly, we focus on reviewing the accuracy or transferability or both of them to show the advantages of adversarial learning, like generative adversarial networks (GANs), in typical computer vision tasks in autonomous systems, including image style transfer, image superresolution, image deblurring/dehazing/rain removal, semantic segmentation, depth estimation, pedestrian detection and person re-identification (re-ID). Then, we further review the performance of RL and meta-learning from the aspects of accuracy or transferability or both of them in autonomous systems, involving pedestrian tracking, robot navigation and robotic manipulation. Finally, we discuss several challenges and future topics for using adversarial learning, RL and meta-learning in autonomous systems.

Chaoqiang Zhao, Qiyu Sun, Chongzhen Zhang et al.

Depth information is important for autonomous systems to perceive environments and estimate their own state. Traditional depth estimation methods, like structure from motion and stereo vision matching, are built on feature correspondences of multiple viewpoints. Meanwhile, the predicted depth maps are sparse. Inferring depth information from a single image (monocular depth estimation) is an ill-posed problem. With the rapid development of deep neural networks, monocular depth estimation based on deep learning has been widely studied recently and achieved promising performance in accuracy. Meanwhile, dense depth maps are estimated from single images by deep neural networks in an end-to-end manner. In order to improve the accuracy of depth estimation, different kinds of network frameworks, loss functions and training strategies are proposed subsequently. Therefore, we survey the current monocular depth estimation methods based on deep learning in this review. Initially, we conclude several widely used datasets and evaluation indicators in deep learning-based depth estimation. Furthermore, we review some representative existing methods according to different training manners: supervised, unsupervised and semi-supervised. Finally, we discuss the challenges and provide some ideas for future researches in monocular depth estimation.

Yang Tang, Chaoqiang Zhao, Jianrui Wang et al.

Autonomous systems possess the features of inferring their own state, understanding their surroundings, and performing autonomous navigation. With the applications of learning systems, like deep learning and reinforcement learning, the visual-based self-state estimation, environment perception and navigation capabilities of autonomous systems have been efficiently addressed, and many new learning-based algorithms have surfaced with respect to autonomous visual perception and navigation. In this review, we focus on the applications of learning-based monocular approaches in ego-motion perception, environment perception and navigation in autonomous systems, which is different from previous reviews that discussed traditional methods. First, we delineate the shortcomings of existing classical visual simultaneous localization and mapping (vSLAM) solutions, which demonstrate the necessity to integrate deep learning techniques. Second, we review the visual-based environmental perception and understanding methods based on deep learning, including deep learning-based monocular depth estimation, monocular ego-motion prediction, image enhancement, object detection, semantic segmentation, and their combinations with traditional vSLAM frameworks. Then, we focus on the visual navigation based on learning systems, mainly including reinforcement learning and deep reinforcement learning. Finally, we examine several challenges and promising directions discussed and concluded in related research of learning systems in the era of computer science and robotics.

Chaoqiang Zhao, Yang Tang, Qiyu Sun et al.

Traditional monocular direct visual odometry (DVO) is one of the most famous methods to estimate the ego-motion of robots and map environments from images simultaneously. However, DVO heavily relies on high-quality images and accurate initial pose estimation during tracking. With the outstanding performance of deep learning, previous works have shown that deep neural networks can effectively learn 6-DoF (Degree of Freedom) poses between frames from monocular image sequences in the unsupervised manner. However, these unsupervised deep learning-based frameworks cannot accurately generate the full trajectory of a long monocular video because of the scale-inconsistency between each pose. To address this problem, we use several geometric constraints to improve the scale-consistency of the pose network, including improving the previous loss function and proposing a novel scale-to-trajectory constraint for unsupervised training. We call the pose network trained by the proposed novel constraint as TrajNet. In addition, a new DVO architecture, called deep direct sparse odometry (DDSO), is proposed to overcome the drawbacks of the previous direct sparse odometry (DSO) framework by embedding TrajNet. Extensive experiments on the KITTI dataset show that the proposed constraints can effectively improve the scale-consistency of TrajNet when compared with previous unsupervised monocular methods, and integration with TrajNet makes the initialization and tracking of DSO more robust and accurate.