Fan Yang

Changho Hwang, Wei Cui, Yifan Xiong et al. · microsoft-research

Sparsely-gated mixture-of-experts (MoE) has been widely adopted to scale deep learning models to trillion-plus parameters with fixed computational cost. The algorithmic performance of MoE relies on its token routing mechanism that forwards each input token to the right sub-models or experts. While token routing dynamically determines the amount of expert workload at runtime, existing systems suffer inefficient computation due to their static execution, namely static parallelism and pipelining, which does not adapt to the dynamic workload. We present Flex, a highly scalable stack design and implementation for MoE with dynamically adaptive parallelism and pipelining. Flex designs an identical layout for distributing MoE model parameters and input data, which can be leveraged by all possible parallelism or pipelining methods without any mathematical inequivalence or tensor migration overhead. This enables adaptive parallelism/pipelining optimization at zero cost during runtime. Based on this key design, Flex also implements various MoE acceleration techniques. Aggregating all techniques, Flex finally delivers huge speedup at any scale -- 4.96x and 5.75x speedup of a single MoE layer over 16 and 2,048 A100 GPUs, respectively, over the previous state-of-the-art. Our evaluation shows that Flex efficiently and effectively runs a real-world MoE-based model named SwinV2-MoE, built upon Swin Transformer V2, a state-of-the-art computer vision architecture. On efficiency, Flex accelerates SwinV2-MoE, achieving up to 1.55x and 2.11x speedup in training and inference over Fairseq, respectively. On effectiveness, the SwinV2-MoE model achieves superior accuracy in both pre-training and down-stream computer vision tasks such as COCO object detection than the counterpart dense model, indicating the readiness of Flex for end-to-end real-world model training and inference.

Yidan Zhang, Ting Zhang, Dong Chen et al. · microsoft-research, pku

While generative modeling has become prevalent across numerous research fields, its integration into the realm of image retrieval remains largely unexplored and underjustified. In this paper, we present a novel methodology, reframing image retrieval as a variant of generative modeling and employing a sequence-to-sequence model. This approach is harmoniously aligned with the current trend towards unification in research, presenting a cohesive framework that allows for end-to-end differentiable searching. This, in turn, facilitates superior performance via direct optimization techniques. The development of our model, dubbed IRGen, addresses the critical technical challenge of converting an image into a concise sequence of semantic units, which is pivotal for enabling efficient and effective search. Extensive experiments demonstrate that our model achieves state-of-the-art performance on three widely-used image retrieval benchmarks as well as two million-scale datasets, yielding significant improvement compared to prior competitive retrieval methods. In addition, the notable surge in precision scores facilitated by generative modeling presents the potential to bypass the reranking phase, which is traditionally indispensable in practical retrieval workflows.

Zhi Chen, Kun Sun, Fan Yang et al.

In this paper, we present a second order spatial compatibility (SC^2) measure based method for efficient and robust point cloud registration (PCR), called SC^2-PCR. Firstly, we propose a second order spatial compatibility (SC^2) measure to compute the similarity between correspondences. It considers the global compatibility instead of local consistency, allowing for more distinctive clustering between inliers and outliers at early stage. Based on this measure, our registration pipeline employs a global spectral technique to find some reliable seeds from the initial correspondences. Then we design a two-stage strategy to expand each seed to a consensus set based on the SC^2 measure matrix. Finally, we feed each consensus set to a weighted SVD algorithm to generate a candidate rigid transformation and select the best model as the final result. Our method can guarantee to find a certain number of outlier-free consensus sets using fewer samplings, making the model estimation more efficient and robust. In addition, the proposed SC^2 measure is general and can be easily plugged into deep learning based frameworks. Extensive experiments are carried out to investigate the performance of our method. Code will be available at \url{https://github.com/ZhiChen902/SC2-PCR}.

Fan Yang

Accurately detecting student behavior in classroom videos can aid in analyzing their classroom performance and improving teaching effectiveness. However, the current accuracy rate in behavior detection is low. To address this challenge, we propose the Student Classroom Behavior Detection method, based on improved YOLOv7. First, we created the Student Classroom Behavior dataset (SCB-Dataset), which includes 18.4k labels and 4.2k images, covering three behaviors: hand raising, reading, and writing. To improve detection accuracy in crowded scenes, we integrated the biformer attention module and Wise-IoU into the YOLOv7 network. Finally, experiments were conducted on the SCB-Dataset, and the model achieved an mAP@0.5 of 79%, resulting in a 1.8% improvement over previous results. The SCB-Dataset and code are available for download at: https://github.com/Whiffe/SCB-dataset.

Zhuyun Zhou, Zongwei Wu, Danda Pani Paudel et al.

Moving object segmentation (MOS) in dynamic scenes is an important, challenging, but under-explored research topic for autonomous driving, especially for sequences obtained from moving ego vehicles. Most segmentation methods leverage motion cues obtained from optical flow maps. However, since these methods are often based on optical flows that are pre-computed from successive RGB frames, this neglects the temporal consideration of events occurring within the inter-frame, consequently constraining its ability to discern objects exhibiting relative staticity but genuinely in motion. To address these limitations, we propose to exploit event cameras for better video understanding, which provide rich motion cues without relying on optical flow. To foster research in this area, we first introduce a novel large-scale dataset called DSEC-MOS for moving object segmentation from moving ego vehicles, which is the first of its kind. For benchmarking, we select various mainstream methods and rigorously evaluate them on our dataset. Subsequently, we devise EmoFormer, a novel network able to exploit the event data. For this purpose, we fuse the event temporal prior with spatial semantic maps to distinguish genuinely moving objects from the static background, adding another level of dense supervision around our object of interest. Our proposed network relies only on event data for training but does not require event input during inference, making it directly comparable to frame-only methods in terms of efficiency and more widely usable in many application cases. The exhaustive comparison highlights a significant performance improvement of our method over all other methods. The source code and dataset are publicly available at: https://github.com/ZZY-Zhou/DSEC-MOS.

Zhuyun Zhou, Zongwei Wu, Rémi Boutteau et al.

Moving Object Detection (MOD) is a critical vision task for successfully achieving safe autonomous driving. Despite plausible results of deep learning methods, most existing approaches are only frame-based and may fail to reach reasonable performance when dealing with dynamic traffic participants. Recent advances in sensor technologies, especially the Event camera, can naturally complement the conventional camera approach to better model moving objects. However, event-based works often adopt a pre-defined time window for event representation, and simply integrate it to estimate image intensities from events, neglecting much of the rich temporal information from the available asynchronous events. Therefore, from a new perspective, we propose RENet, a novel RGB-Event fusion Network, that jointly exploits the two complementary modalities to achieve more robust MOD under challenging scenarios for autonomous driving. Specifically, we first design a temporal multi-scale aggregation module to fully leverage event frames from both the RGB exposure time and larger intervals. Then we introduce a bi-directional fusion module to attentively calibrate and fuse multi-modal features. To evaluate the performance of our network, we carefully select and annotate a sub-MOD dataset from the commonly used DSEC dataset. Extensive experiments demonstrate that our proposed method performs significantly better than the state-of-the-art RGB-Event fusion alternatives. The source code and dataset are publicly available at: https://github.com/ZZY-Zhou/RENet.

Zhiyang Chen, Yousong Zhu, Zhaowen Li et al.

Visual tasks vary a lot in their output formats and concerned contents, therefore it is hard to process them with an identical structure. One main obstacle lies in the high-dimensional outputs in object-level visual tasks. In this paper, we propose an object-centric vision framework, Obj2Seq. Obj2Seq takes objects as basic units, and regards most object-level visual tasks as sequence generation problems of objects. Therefore, these visual tasks can be decoupled into two steps. First recognize objects of given categories, and then generate a sequence for each of these objects. The definition of the output sequences varies for different tasks, and the model is supervised by matching these sequences with ground-truth targets. Obj2Seq is able to flexibly determine input categories to satisfy customized requirements, and be easily extended to different visual tasks. When experimenting on MS COCO, Obj2Seq achieves 45.7% AP on object detection, 89.0% AP on multi-label classification and 65.0% AP on human pose estimation. These results demonstrate its potential to be generally applied to different visual tasks. Code has been made available at: https://github.com/CASIA-IVA-Lab/Obj2Seq.

Jiawei Yang, Hanbo Chen, Yu Zhao et al.

Whole slide image (WSI) classification often relies on deep weakly supervised multiple instance learning (MIL) methods to handle gigapixel resolution images and slide-level labels. Yet the decent performance of deep learning comes from harnessing massive datasets and diverse samples, urging the need for efficient training pipelines for scaling to large datasets and data augmentation techniques for diversifying samples. However, current MIL-based WSI classification pipelines are memory-expensive and computation-inefficient since they usually assemble tens of thousands of patches as bags for computation. On the other hand, despite their popularity in other tasks, data augmentations are unexplored for WSI MIL frameworks. To address them, we propose ReMix, a general and efficient framework for MIL based WSI classification. It comprises two steps: reduce and mix. First, it reduces the number of instances in WSI bags by substituting instances with instance prototypes, i.e., patch cluster centroids. Then, we propose a ``Mix-the-bag'' augmentation that contains four online, stochastic and flexible latent space augmentations. It brings diverse and reliable class-identity-preserving semantic changes in the latent space while enforcing semantic-perturbation invariance. We evaluate ReMix on two public datasets with two state-of-the-art MIL methods. In our experiments, consistent improvements in precision, accuracy, and recall have been achieved but with orders of magnitude reduced training time and memory consumption, demonstrating ReMix's effectiveness and efficiency. Code is available.

Zhiqi Lin, Youshan Miao, Guanbin Xu et al. · microsoft-research

Increasingly complex and diverse deep neural network (DNN) models necessitate distributing the execution across multiple devices for training and inference tasks, and also require carefully planned schedules for performance. However, existing practices often rely on predefined schedules that may not fully exploit the benefits of emerging diverse model-aware operator placement strategies. Handcrafting high-efficiency schedules can be challenging due to the large and varying schedule space. This paper presents Tessel, an automated system that searches for efficient schedules for distributed DNN training and inference for diverse operator placement strategies. To reduce search costs, Tessel leverages the insight that the most efficient schedules often exhibit repetitive pattern (repetend) across different data inputs. This leads to a two-phase approach: repetend construction and schedule completion. By exploring schedules for various operator placement strategies, Tessel significantly improves both training and inference performance. Experiments with representative DNN models demonstrate that Tessel achieves up to 5.5x training performance speedup and up to 38% inference latency reduction.

Zhiqi Lin, Youshan Miao, Guodong Liu et al. · microsoft-research

With the growing model size, deep neural networks (DNN) are increasingly trained over massive GPU accelerators, which demands a proper parallelization plan that transforms a DNN model into fine-grained tasks and then schedules them to GPUs for execution. Due to the large search space, the contemporary parallelization plan generators often rely on empirical rules that couple transformation and scheduling, and fall short in exploring more flexible schedules that yield better memory usage and compute efficiency. This tension can be exacerbated by the emerging models with increasing complexity in their structure and model size. SuperScaler is a system that facilitates the design and generation of highly flexible parallelization plans. It formulates the plan design and generation into three sequential phases explicitly: model transformation, space-time scheduling, and data dependency preserving. Such a principled approach decouples multiple seemingly intertwined factors and enables the composition of highly flexible parallelization plans. As a result, SuperScaler can not only generate empirical parallelization plans, but also construct new plans that achieve up to 3.5X speedup compared to state-of-the-art solutions like DeepSpeed, Megatron and Alpa, for emerging DNN models like Swin-Transformer and AlphaFold2, as well as well-optimized models like GPT-3.

Cong Guo, Yuxian Qiu, Jingwen Leng et al. · microsoft-research

An activation function is an element-wise mathematical function and plays a crucial role in deep neural networks (DNN). Many novel and sophisticated activation functions have been proposed to improve the DNN accuracy but also consume massive memory in the training process with back-propagation. In this study, we propose the nested forward automatic differentiation (Forward-AD), specifically for the element-wise activation function for memory-efficient DNN training. We deploy nested Forward-AD in two widely-used deep learning frameworks, TensorFlow and PyTorch, which support the static and dynamic computation graph, respectively. Our evaluation shows that nested Forward-AD reduces the memory footprint by up to 1.97x than the baseline model and outperforms the recomputation by 20% under the same memory reduction ratio.

Zhichao Lu, Ran Cheng, Shihua Huang et al.

The architectural advancements in deep neural networks have led to remarkable leap-forwards across a broad array of computer vision tasks. Instead of relying on human expertise, neural architecture search (NAS) has emerged as a promising avenue toward automating the design of architectures. While recent achievements in image classification have suggested opportunities, the promises of NAS have yet to be thoroughly assessed on more challenging tasks of semantic segmentation. The main challenges of applying NAS to semantic segmentation arise from two aspects: (i) high-resolution images to be processed; (ii) additional requirement of real-time inference speed (i.e., real-time semantic segmentation) for applications such as autonomous driving. To meet such challenges, we propose a surrogate-assisted multi-objective method in this paper. Through a series of customized prediction models, our method effectively transforms the original NAS task into an ordinary multi-objective optimization problem. Followed by a hierarchical pre-screening criterion for in-fill selection, our method progressively achieves a set of efficient architectures trading-off between segmentation accuracy and inference speed. Empirical evaluations on three benchmark datasets together with an application using Huawei Atlas 200 DK suggest that our method can identify architectures significantly outperforming existing state-of-the-art architectures designed both manually by human experts and automatically by other NAS methods.

Zelong Zeng, Zheng Wang, Fan Yang et al.

The large variation of viewpoint and irrelevant content around the target always hinder accurate image retrieval and its subsequent tasks. In this paper, we investigate an extremely challenging task: given a ground-view image of a landmark, we aim to achieve cross-view geo-localization by searching out its corresponding satellite-view images. Specifically, the challenge comes from the gap between ground-view and satellite-view, which includes not only large viewpoint changes (some parts of the landmark may be invisible from front view to top view) but also highly irrelevant background (the target landmark tend to be hidden in other surrounding buildings), making it difficult to learn a common representation or a suitable mapping. To address this issue, we take advantage of drone-view information as a bridge between ground-view and satellite-view domains. We propose a Peer Learning and Cross Diffusion (PLCD) framework. PLCD consists of three parts: 1) a peer learning across ground-view and drone-view to find visible parts to benefit ground-drone cross-view representation learning; 2) a patch-based network for satellite-drone cross-view representation learning; 3) a cross diffusion between ground-drone space and satellite-drone space. Extensive experiments conducted on the University-Earth and University-Google datasets show that our method outperforms state-of-the-arts significantly.

Zelong Zeng, Fan Yang, Hong Liu et al.

Recent deep metric learning (DML) methods typically leverage solely class labels to keep positive samples far away from negative ones. However, this type of method normally ignores the crucial knowledge hidden in the data (e.g., intra-class information variation), which is harmful to the generalization of the trained model. To alleviate this problem, in this paper we propose Online Batch Diffusion-based Self-Distillation (OBD-SD) for DML. Specifically, we first propose a simple but effective Progressive Self-Distillation (PSD), which distills the knowledge progressively from the model itself during training. The soft distance targets achieved by PSD can present richer relational information among samples, which is beneficial for the diversity of embedding representations. Then, we extend PSD with an Online Batch Diffusion Process (OBDP), which is to capture the local geometric structure of manifolds in each batch, so that it can reveal the intrinsic relationships among samples in the batch and produce better soft distance targets. Note that our OBDP is able to restore the insufficient manifold relationships obtained by the original PSD and achieve significant performance improvement. Our OBD-SD is a flexible framework that can be integrated into state-of-the-art (SOTA) DML methods. Extensive experiments on various benchmarks, namely CUB200, CARS196, and Stanford Online Products, demonstrate that our OBD-SD consistently improves the performance of the existing DML methods on multiple datasets with negligible additional training time, achieving very competitive results. Code: \url{https://github.com/ZelongZeng/OBD-SD_Pytorch}

Jiaxin Deng, Dong Shen, Haojie Pan et al.

Video understanding is an important task in short video business platforms and it has a wide application in video recommendation and classification. Most of the existing video understanding works only focus on the information that appeared within the video content, including the video frames, audio and text. However, introducing common sense knowledge from the external Knowledge Graph (KG) dataset is essential for video understanding when referring to the content which is less relevant to the video. Owing to the lack of video knowledge graph dataset, the work which integrates video understanding and KG is rare. In this paper, we propose a heterogeneous dataset that contains the multi-modal video entity and fruitful common sense relations. This dataset also provides multiple novel video inference tasks like the Video-Relation-Tag (VRT) and Video-Relation-Video (VRV) tasks. Furthermore, based on this dataset, we propose an end-to-end model that jointly optimizes the video understanding objective with knowledge graph embedding, which can not only better inject factual knowledge into video understanding but also generate effective multi-modal entity embedding for KG. Comprehensive experiments indicate that combining video understanding embedding with factual knowledge benefits the content-based video retrieval performance. Moreover, it also helps the model generate better knowledge graph embedding which outperforms traditional KGE-based methods on VRT and VRV tasks with at least 42.36% and 17.73% improvement in HITS@10.

Xing Cheng, Xiangyu Wu, Dong Shen et al.

Video grounding aims to locate the timestamps best matching the query description within an untrimmed video. Prevalent methods can be divided into moment-level and clip-level frameworks. Moment-level approaches directly predict the probability of each transient moment to be the boundary in a global perspective, and they usually perform better in coarse grounding. On the other hand, clip-level ones aggregate the moments in different time windows into proposals and then deduce the most similar one, leading to its advantage in fine-grained grounding. In this paper, we propose a multi-level unified framework to enhance performance by leveraging the merits of both moment-level and clip-level methods. Moreover, a novel generation-guided paradigm in both levels is adopted. It introduces a multi-modal generator to produce the implicit boundary feature and clip feature, later regarded as queries to calculate the boundary scores by a discriminator. The generation-guided solution enhances video grounding from a two-unique-modals' match task to a cross-modal attention task, which steps out of the previous framework and obtains notable gains. The proposed Generation-guided Multi-level Unified network (GMU) surpasses previous methods and reaches State-Of-The-Art on various benchmarks with disparate features, e.g., Charades-STA, ActivityNet captions.

Guoqiang Jin, Fan Yang, Mingshan Sun et al.

Self-supervised pre-training and transformer-based networks have significantly improved the performance of object detection. However, most of the current self-supervised object detection methods are built on convolutional-based architectures. We believe that the transformers' sequence characteristics should be considered when designing a transformer-based self-supervised method for the object detection task. To this end, we propose SeqCo-DETR, a novel Sequence Consistency-based self-supervised method for object DEtection with TRansformers. SeqCo-DETR defines a simple but effective pretext by minimizes the discrepancy of the output sequences of transformers with different image views as input and leverages bipartite matching to find the most relevant sequence pairs to improve the sequence-level self-supervised representation learning performance. Furthermore, we provide a mask-based augmentation strategy incorporated with the sequence consistency strategy to extract more representative contextual information about the object for the object detection task. Our method achieves state-of-the-art results on MS COCO (45.8 AP) and PASCAL VOC (64.1 AP), demonstrating the effectiveness of our approach.

Zihao Lin, Jinrong Li, Fan Yang et al.

Offline Handwritten Mathematical Expression Recognition (HMER) has been dramatically advanced recently by employing tree decoders as part of the encoder-decoder method. Despite the tree decoder-based methods regard the expressions as a tree and parse 2D spatial structure to the tree nodes sequence, the performance of existing works is still poor due to the inevitable tree nodes prediction errors. Besides, they lack syntax rules to regulate the output of expressions. In this paper, we propose a novel model called Spatial Attention and Syntax Rule Enhanced Tree Decoder (SS-TD), which is equipped with spatial attention mechanism to alleviate the prediction error of tree structure and use syntax masks (obtained from the transformation of syntax rules) to constrain the occurrence of ungrammatical mathematical expression. In this way, our model can effectively describe tree structure and increase the accuracy of output expression. Experiments show that SS-TD achieves better recognition performance than prior models on CROHME 14/16/19 datasets, demonstrating the effectiveness of our model.

Xiaowen Shi, Ze Wang, Yuanying Cai et al.

Nowadays, the mainstream approach in position allocation system is to utilize a reinforcement learning model to allocate appropriate locations for items in various channels and then mix them into the feed. There are two types of data employed to train reinforcement learning (RL) model for position allocation, named strategy data and random data. Strategy data is collected from the current online model, it suffers from an imbalanced distribution of state-action pairs, resulting in severe overestimation problems during training. On the other hand, random data offers a more uniform distribution of state-action pairs, but is challenging to obtain in industrial scenarios as it could negatively impact platform revenue and user experience due to random exploration. As the two types of data have different distributions, designing an effective strategy to leverage both types of data to enhance the efficacy of the RL model training has become a highly challenging problem. In this study, we propose a framework named Multi-Distribution Data Learning (MDDL) to address the challenge of effectively utilizing both strategy and random data for training RL models on mixed multi-distribution data. Specifically, MDDL incorporates a novel imitation learning signal to mitigate overestimation problems in strategy data and maximizes the RL signal for random data to facilitate effective learning. In our experiments, we evaluated the proposed MDDL framework in a real-world position allocation system and demonstrated its superior performance compared to the previous baseline. MDDL has been fully deployed on the Meituan food delivery platform and currently serves over 300 million users.

Xiaowen Ma, Shuning Ge, Fan Yang et al.

Transformer-based architectures dominate time series modeling by enabling global attention over all timestamps, yet their rigid 'one-size-fits-all' context aggregation fails to address two critical challenges in real-world data: (1) inherent lag effects, where the relevance of historical timestamps to a query varies dynamically; (2) anomalous segments, which introduce noisy signals that degrade forecasting accuracy. To resolve these problems, we propose the Temporal Mix of Experts (TMOE), a novel attention-level mechanism that reimagines key-value (K-V) pairs as local experts (each specialized in a distinct temporal context) and performs adaptive expert selection for each query via localized filtering of irrelevant timestamps. Complementing this local adaptation, a shared global expert preserves the Transformer's strength in capturing long-range dependencies. We then replace the vanilla attention mechanism in popular time-series Transformer frameworks (i.e., PatchTST and Timer) with TMOE, without extra structural modifications, yielding our specific version TimeExpert and general version TimeExpert-G. Extensive experiments on seven real-world long-term forecasting benchmarks demonstrate that TimeExpert and TimeExpert-G outperform state-of-the-art methods. Code is available at https://github.com/xwmaxwma/TimeExpert.

Manthan Patel, Fan Yang, Yuheng Qiu et al.

We present TartanGround, a large-scale, multi-modal dataset to advance the perception and autonomy of ground robots operating in diverse environments. This dataset, collected in various photorealistic simulation environments includes multiple RGB stereo cameras for 360-degree coverage, along with depth, optical flow, stereo disparity, LiDAR point clouds, ground truth poses, semantic segmented images, and occupancy maps with semantic labels. Data is collected using an integrated automatic pipeline, which generates trajectories mimicking the motion patterns of various ground robot platforms, including wheeled and legged robots. We collect 910 trajectories across 70 environments, resulting in 1.5 million samples. Evaluations on occupancy prediction and SLAM tasks reveal that state-of-the-art methods trained on existing datasets struggle to generalize across diverse scenes. TartanGround can serve as a testbed for training and evaluation of a broad range of learning-based tasks, including occupancy prediction, SLAM, neural scene representation, perception-based navigation, and more, enabling advancements in robotic perception and autonomy towards achieving robust models generalizable to more diverse scenarios. The dataset and codebase are available on the webpage: https://tartanair.org/tartanground

Yijia Zhang, Yibo Han, Shijie Cao et al.

Running out of GPU memory has become a main bottleneck for large-scale DNN training. How to reduce the memory footprint during training has received intensive research attention. We find that previous gradient accumulation reduces activation memory but fails to be compatible with gradient memory reduction due to a contradiction between preserving gradients and releasing gradients. To address this issue, we propose a novel optimizer accumulation method for Adam, named Adam Accumulation (AdamA), which enables reducing both activation and gradient memory. Specifically, AdamA directly integrates gradients into optimizer states and accumulates optimizer states over micro-batches, so that gradients can be released immediately after use. We mathematically and experimentally demonstrate AdamA yields the same convergence properties as Adam. Evaluated on transformer-based models, AdamA achieves up to 23% memory reduction compared to gradient accumulation with less than 2% degradation in training throughput. Notably, AdamA can work together with memory reduction methods for optimizer states to fit 1.26x~3.14x larger models over PyTorch and DeepSpeed baseline on GPUs with different memory capacities.

Yazhong Si, Fan Yang, Ya Guo et al.

Numerous sand dust image enhancement algorithms have been proposed in recent years. To our best acknowledge, however, most methods evaluated their performance with no-reference way using few selected real-world images from internet. It is unclear how to quantitatively analysis the performance of the algorithms in a supervised way and how we could gauge the progress in the field. Moreover, due to the absence of large-scale benchmark datasets, there are no well-known reports of data-driven based method for sand dust image enhancement up till now. To advance the development of deep learning-based algorithms for sand dust image reconstruction, while enabling supervised objective evaluation of algorithm performance. In this paper, we presented a comprehensive perceptual study and analysis of real-world sand dust images, then constructed a Sand-dust Image Reconstruction Benchmark (SIRB) for training Convolutional Neural Networks (CNNs) and evaluating algorithms performance. In addition, we adopted the existing image transformation neural network trained on SIRB as baseline to illustrate the generalization of SIRB for training CNNs. Finally, we conducted the qualitative and quantitative evaluation to demonstrate the performance and limitations of the state-of-the-arts (SOTA), which shed light on future research in sand dust image reconstruction.

Wuyang Chen, Xianzhi Du, Fan Yang et al.

This work presents a simple vision transformer design as a strong baseline for object localization and instance segmentation tasks. Transformers recently demonstrate competitive performance in image classification tasks. To adopt ViT to object detection and dense prediction tasks, many works inherit the multistage design from convolutional networks and highly customized ViT architectures. Behind this design, the goal is to pursue a better trade-off between computational cost and effective aggregation of multiscale global contexts. However, existing works adopt the multistage architectural design as a black-box solution without a clear understanding of its true benefits. In this paper, we comprehensively study three architecture design choices on ViT -- spatial reduction, doubled channels, and multiscale features -- and demonstrate that a vanilla ViT architecture can fulfill this goal without handcrafting multiscale features, maintaining the original ViT design philosophy. We further complete a scaling rule to optimize our model's trade-off on accuracy and computation cost / model size. By leveraging a constant feature resolution and hidden size throughout the encoder blocks, we propose a simple and compact ViT architecture called Universal Vision Transformer (UViT) that achieves strong performance on COCO object detection and instance segmentation tasks.

Jimiama M. Mase, Natalie Leesakul, Fan Yang et al.

Automatically understanding and recognising human affective states using images and computer vision can improve human-computer and human-robot interaction. However, privacy has become an issue of great concern, as the identities of people used to train affective models can be exposed in the process. For instance, malicious individuals could exploit images from users and assume their identities. In addition, affect recognition using images can lead to discriminatory and algorithmic bias, as certain information such as race, gender, and age could be assumed based on facial features. Possible solutions to protect the privacy of users and avoid misuse of their identities are to: (1) extract anonymised facial features, namely action units (AU) from a database of images, discard the images and use AUs for processing and training, and (2) federated learning (FL) i.e. process raw images in users' local machines (local processing) and send the locally trained models to the main processing machine for aggregation (central processing). In this paper, we propose a two-level deep learning architecture for affect recognition that uses AUs in level 1 and FL in level 2 to protect users' identities. The architecture consists of recurrent neural networks to capture the temporal relationships amongst the features and predict valence and arousal affective states. In our experiments, we evaluate the performance of our privacy-preserving architecture using different variations of recurrent neural networks on RECOLA, a comprehensive multimodal affective database. Our results show state-of-the-art performance of $0.426$ for valence and $0.401$ for arousal using the Concordance Correlation Coefficient evaluation metric, demonstrating the feasibility of developing models for affect recognition that are both accurate and ensure privacy.

Fan Yang, Chao Cao, Hongbiao Zhu et al.

The problem of path planning in unknown environments remains a challenging problem - as the environment is gradually observed during the navigation, the underlying planner has to update the environment representation and replan, promptly and constantly, to account for the new observations. In this paper, we present a visibility graph-based planning framework capable of dealing with navigation tasks in both known and unknown environments. The planner employs a polygonal representation of the environment and constructs the representation by extracting edge points around obstacles to form enclosed polygons. With that, the method dynamically updates a global visibility graph using a two-layered data structure, expanding the visibility edges along with the navigation and removing edges that become occluded by newly observed obstacles. When navigating in unknown environments, the method is attemptable in discovering a way to the goal by picking up the environment layout on the fly, updating the visibility graph, and fast re-planning corresponding to the newly observed environment. We evaluate the method in simulated and real-world settings. The method shows the capability to attempt and navigate through unknown environments, reducing the travel time by up to 12-47% from search-based methods: A*, D* Lite, and more than 24-35% than sampling-based methods: RRT*, BIT*, and SPARS.

Qiang Zhai, Xin Li, Fan Yang et al.

Automatically detecting/segmenting object(s) that blend in with their surroundings is difficult for current models. A major challenge is that the intrinsic similarities between such foreground objects and background surroundings make the features extracted by deep model indistinguishable. To overcome this challenge, an ideal model should be able to seek valuable, extra clues from the given scene and incorporate them into a joint learning framework for representation co-enhancement. With this inspiration, we design a novel Mutual Graph Learning (MGL) model, which generalizes the idea of conventional mutual learning from regular grids to the graph domain. Specifically, MGL decouples an image into two task-specific feature maps -- one for roughly locating the target and the other for accurately capturing its boundary details -- and fully exploits the mutual benefits by recurrently reasoning their high-order relations through graphs. Importantly, in contrast to most mutual learning approaches that use a shared function to model all between-task interactions, MGL is equipped with typed functions for handling different complementary relations to maximize information interactions. Experiments on challenging datasets, including CHAMELEON, CAMO and COD10K, demonstrate the effectiveness of our MGL with superior performance to existing state-of-the-art methods.

Ao Luo, Xin Li, Fan Yang et al.

In this paper, we study the problem of salient object detection (SOD) for RGB-D images using both color and depth information.A major technical challenge in performing salient object detection fromRGB-D images is how to fully leverage the two complementary data sources. Current works either simply distill prior knowledge from the corresponding depth map for handling the RGB-image or blindly fuse color and geometric information to generate the coarse depth-aware representations, hindering the performance of RGB-D saliency detectors.In this work, we introduceCascade Graph Neural Networks(Cas-Gnn),a unified framework which is capable of comprehensively distilling and reasoning the mutual benefits between these two data sources through a set of cascade graphs, to learn powerful representations for RGB-D salient object detection. Cas-Gnn processes the two data sources individually and employs a novelCascade Graph Reasoning(CGR) module to learn powerful dense feature embeddings, from which the saliency map can be easily inferred. Contrast to the previous approaches, the explicitly modeling and reasoning of high-level relations between complementary data sources allows us to better overcome challenges such as occlusions and ambiguities. Extensive experiments demonstrate that Cas-Gnn achieves significantly better performance than all existing RGB-DSOD approaches on several widely-used benchmarks.

Fan Yang, Xiao Xiao

Blur detection is the separation of blurred and clear regions of an image, which is an important and challenging task in computer vision. In this work, we regard blur detection as an image segmentation problem. Inspired by the success of the U-net architecture for image segmentation, we design a Multi-Scale Dilated convolutional neural network based on U-net, which we call MSDU-net. The MSDU-net uses a group of multi-scale feature extractors with dilated convolutions to extract texture information at different scales. The U-shape architecture of the MSDU-net fuses the different-scale texture features and generates a semantic feature which allows us to achieve better results on the blur detection task. We show that using the MSDU-net we are able to outperform other state of the art blur detection methods on two publicly available benchmarks.

Fan Yang, Cheng Lu, Yandong Guo et al.

Feature pyramid architecture has been broadly adopted in object detection and segmentation to deal with multi-scale problem. However, in this paper we show that the capacity of the architecture has not been fully explored due to the inadequate utilization of the supervision information. Such insufficient utilization is caused by the supervision signal degradation in back propagation. Thus inspired, we propose a dually supervised method, named dually supervised FPN (DSFPN), to enhance the supervision signal when training the feature pyramid network (FPN). In particular, DSFPN is constructed by attaching extra prediction (i.e., detection or segmentation) heads to the bottom-up subnet of FPN. Hence, the features can be optimized by the additional heads before being forwarded to subsequent networks. Further, the auxiliary heads can serve as a regularization term to facilitate the model training. In addition, to strengthen the capability of the detection heads in DSFPN for handling two inhomogeneous tasks, i.e., classification and regression, the originally shared hidden feature space is separated by decoupling classification and regression subnets. To demonstrate the generalizability, effectiveness, and efficiency of the proposed method, DSFPN is integrated into four representative detectors (Faster RCNN, Mask RCNN, Cascade RCNN, and Cascade Mask RCNN) and assessed on the MS COCO dataset. Promising precision improvement, state-of-the-art performance, and negligible additional computational cost are demonstrated through extensive experiments. Code will be provided.

Fan Yang, Ryota Hinami, Yusuke Matsui et al.

Diffusion is commonly used as a ranking or re-ranking method in retrieval tasks to achieve higher retrieval performance, and has attracted lots of attention in recent years. A downside to diffusion is that it performs slowly in comparison to the naive k-NN search, which causes a non-trivial online computational cost on large datasets. To overcome this weakness, we propose a novel diffusion technique in this paper. In our work, instead of applying diffusion to the query, we pre-compute the diffusion results of each element in the database, making the online search a simple linear combination on top of the k-NN search process. Our proposed method becomes 10~ times faster in terms of online search speed. Moreover, we propose to use late truncation instead of early truncation in previous works to achieve better retrieval performance.

Mahyar Najibi, Fan Yang, Qiaosong Wang et al.

In this work, we propose an efficient and effective approach for unconstrained salient object detection in images using deep convolutional neural networks. Instead of generating thousands of candidate bounding boxes and refining them, our network directly learns to generate the saliency map containing the exact number of salient objects. During training, we convert the ground-truth rectangular boxes to Gaussian distributions that better capture the ROI regarding individual salient objects. During inference, the network predicts Gaussian distributions centered at salient objects with an appropriate covariance, from which bounding boxes are easily inferred. Notably, our network performs saliency map prediction without pixel-level annotations, salient object detection without object proposals, and salient object subitizing simultaneously, all in a single pass within a unified framework. Extensive experiments show that our approach outperforms existing methods on various datasets by a large margin, and achieves more than 100 fps with VGG16 network on a single GPU during inference.

Joe Yue-Hei Ng, Fan Yang, Larry S. Davis

Deep convolutional neural networks have been successfully applied to image classification tasks. When these same networks have been applied to image retrieval, the assumption has been made that the last layers would give the best performance, as they do in classification. We show that for instance-level image retrieval, lower layers often perform better than the last layers in convolutional neural networks. We present an approach for extracting convolutional features from different layers of the networks, and adopt VLAD encoding to encode features into a single vector for each image. We investigate the effect of different layers and scales of input images on the performance of convolutional features using the recent deep networks OxfordNet and GoogLeNet. Experiments demonstrate that intermediate layers or higher layers with finer scales produce better results for image retrieval, compared to the last layer. When using compressed 128-D VLAD descriptors, our method obtains state-of-the-art results and outperforms other VLAD and CNN based approaches on two out of three test datasets. Our work provides guidance for transferring deep networks trained on image classification to image retrieval tasks.