Wei Zhang

Peng Gao, Jiaming Han, Renrui Zhang et al. · berkeley, stanford

How to efficiently transform large language models (LLMs) into instruction followers is recently a popular research direction, while training LLM for multi-modal reasoning remains less explored. Although the recent LLaMA-Adapter demonstrates the potential to handle visual inputs with LLMs, it still cannot generalize well to open-ended visual instructions and lags behind GPT-4. In this paper, we present LLaMA-Adapter V2, a parameter-efficient visual instruction model. Specifically, we first augment LLaMA-Adapter by unlocking more learnable parameters (e.g., norm, bias and scale), which distribute the instruction-following ability across the entire LLaMA model besides adapters. Secondly, we propose an early fusion strategy to feed visual tokens only into the early LLM layers, contributing to better visual knowledge incorporation. Thirdly, a joint training paradigm of image-text pairs and instruction-following data is introduced by optimizing disjoint groups of learnable parameters. This strategy effectively alleviates the interference between the two tasks of image-text alignment and instruction following and achieves strong multi-modal reasoning with only a small-scale image-text and instruction dataset. During inference, we incorporate additional expert models (e.g. captioning/OCR systems) into LLaMA-Adapter to further enhance its image understanding capability without incurring training costs. Compared to the original LLaMA-Adapter, our LLaMA-Adapter V2 can perform open-ended multi-modal instructions by merely introducing 14M parameters over LLaMA. The newly designed framework also exhibits stronger language-only instruction-following capabilities and even excels in chat interactions. Our code and models are available at https://github.com/ZrrSkywalker/LLaMA-Adapter.

Runmin Cong, Wenyu Yang, Wei Zhang et al.

Due to the light absorption and scattering induced by the water medium, underwater images usually suffer from some degradation problems, such as low contrast, color distortion, and blurring details, which aggravate the difficulty of downstream underwater understanding tasks. Therefore, how to obtain clear and visually pleasant images has become a common concern of people, and the task of underwater image enhancement (UIE) has also emerged as the times require. Among existing UIE methods, Generative Adversarial Networks (GANs) based methods perform well in visual aesthetics, while the physical model-based methods have better scene adaptability. Inheriting the advantages of the above two types of models, we propose a physical model-guided GAN model for UIE in this paper, referred to as PUGAN. The entire network is under the GAN architecture. On the one hand, we design a Parameters Estimation subnetwork (Par-subnet) to learn the parameters for physical model inversion, and use the generated color enhancement image as auxiliary information for the Two-Stream Interaction Enhancement sub-network (TSIE-subnet). Meanwhile, we design a Degradation Quantization (DQ) module in TSIE-subnet to quantize scene degradation, thereby achieving reinforcing enhancement of key regions. On the other hand, we design the Dual-Discriminators for the style-content adversarial constraint, promoting the authenticity and visual aesthetics of the results. Extensive experiments on three benchmark datasets demonstrate that our PUGAN outperforms state-of-the-art methods in both qualitative and quantitative metrics.

Jincheng Lu, Xipeng Yang, Jin Ye et al.

Multi-Camera Multi-Target Tracking (MCMT) is a computer vision technique that involves tracking multiple targets simultaneously across multiple cameras. MCMT in urban traffic visual analysis faces great challenges due to the complex and dynamic nature of urban traffic scenes, where multiple cameras with different views and perspectives are often used to cover a large city-scale area. Targets in urban traffic scenes often undergo occlusion, illumination changes, and perspective changes, making it difficult to associate targets across different cameras accurately. To overcome these challenges, we propose a novel systematic MCMT framework, called CityTrack. Specifically, we present a Location-Aware SCMT tracker which integrates various advanced techniques to improve its effectiveness in the MCMT task and propose a novel Box-Grained Matching (BGM) method for the ICA module to solve the aforementioned problems. We evaluated our approach on the public test set of the CityFlowV2 dataset and achieved an IDF1 of 84.91%, ranking 1st in the 2022 AI CITY CHALLENGE. Our experimental results demonstrate the effectiveness of our approach in overcoming the challenges posed by urban traffic scenes.

Yu Zhang, Hao Zeng, Bowen Ma et al.

This work proposes a novel face-swapping framework FlowFace++, utilizing explicit semantic flow supervision and end-to-end architecture to facilitate shape-aware face-swapping. Specifically, our work pretrains a facial shape discriminator to supervise the face swapping network. The discriminator is shape-aware and relies on a semantic flow-guided operation to explicitly calculate the shape discrepancies between the target and source faces, thus optimizing the face swapping network to generate highly realistic results. The face swapping network is a stack of a pre-trained face-masked autoencoder (MAE), a cross-attention fusion module, and a convolutional decoder. The MAE provides a fine-grained facial image representation space, which is unified for the target and source faces and thus facilitates final realistic results. The cross-attention fusion module carries out the source-to-target face swapping in a fine-grained latent space while preserving other attributes of the target image (e.g. expression, head pose, hair, background, illumination, etc). Lastly, the convolutional decoder further synthesizes the swapping results according to the face-swapping latent embedding from the cross-attention fusion module. Extensive quantitative and qualitative experiments on in-the-wild faces demonstrate that our FlowFace++ outperforms the state-of-the-art significantly, particularly while the source face is obstructed by uneven lighting or angle offset.

Bohui Shen, Wei Zhang, Xubiao Liu et al.

Positron emission tomography (PET) serves as an essential tool for diagnosis of encephalopathy and brain science research. However, it suffers from the limited choice of tracers. Nowadays, with the wide application of PET imaging in neuropsychiatric treatment, 6-18F-fluoro-3, 4-dihydroxy-L-phenylalanine (DOPA) has been found to be more effective than 18F-labeled fluorine-2-deoxyglucose (FDG) in the field. Nevertheless, due to the complexity of its preparation and other limitations, DOPA is far less widely used than FDG. To address this issue, a tracer conversion invertible neural network (TC-INN) for image projection is developed to map FDG images to DOPA images through deep learning. More diagnostic information is obtained by generating PET images from FDG to DOPA. Specifically, the proposed TC-INN consists of two separate phases, one for training traceable data, the other for rebuilding new data. The reference DOPA PET image is used as a learning target for the corresponding network during the training process of tracer conversion. Meanwhile, the invertible network iteratively estimates the resultant DOPA PET data and compares it to the reference DOPA PET data. Notably, the reversible model employs variable enhancement technique to achieve better power generation. Moreover, image registration needs to be performed before training due to the angular deviation of the acquired FDG and DOPA data information. Experimental results exhibited excellent generation capability in mapping between FDG and DOPA, suggesting that PET tracer conversion has great potential in the case of limited tracer applications.

Pulin Li, Guocheng Wu, Li Yin et al.

Social manufacturing leverages community collaboration and scattered resources to realize mass individualization in modern industry. However, this paradigm shift also introduces substantial challenges in quality control, particularly in defect detection. The main difficulties stem from three aspects. First, products often have highly customized configurations. Second, production typically involves fragmented, small-batch orders. Third, imaging environments vary considerably across distributed sites. To overcome the scarcity of real-world datasets and tailored algorithms, we introduce the Mass Individualization Robust Anomaly Detection (MIRAD) dataset. As the first benchmark explicitly designed for anomaly detection in social manufacturing, MIRAD captures three critical dimensions of this domain: (1) diverse individualized products with large intra-class variation, (2) data collected from six geographically dispersed manufacturing nodes, and (3) substantial imaging heterogeneity, including variations in lighting, background, and motion conditions. We then conduct extensive evaluations of state-of-the-art (SOTA) anomaly detection methods on MIRAD, covering one-class, multi-class, and zero-shot approaches. Results show a significant performance drop across all models compared with conventional benchmarks, highlighting the unresolved complexities of defect detection in real-world individualized production. By bridging industrial requirements and academic research, MIRAD provides a realistic foundation for developing robust quality control solutions essential for Industry 5.0. The dataset is publicly available at https://github.com/wu33learn/MIRAD.

Renrui Zhang, Ziyu Guo, Wei Zhang et al.

Recently, zero-shot and few-shot learning via Contrastive Vision-Language Pre-training (CLIP) have shown inspirational performance on 2D visual recognition, which learns to match images with their corresponding texts in open-vocabulary settings. However, it remains under explored that whether CLIP, pre-trained by large-scale image-text pairs in 2D, can be generalized to 3D recognition. In this paper, we identify such a setting is feasible by proposing PointCLIP, which conducts alignment between CLIP-encoded point cloud and 3D category texts. Specifically, we encode a point cloud by projecting it into multi-view depth maps without rendering, and aggregate the view-wise zero-shot prediction to achieve knowledge transfer from 2D to 3D. On top of that, we design an inter-view adapter to better extract the global feature and adaptively fuse the few-shot knowledge learned from 3D into CLIP pre-trained in 2D. By just fine-tuning the lightweight adapter in the few-shot settings, the performance of PointCLIP could be largely improved. In addition, we observe the complementary property between PointCLIP and classical 3D-supervised networks. By simple ensembling, PointCLIP boosts baseline's performance and even surpasses state-of-the-art models. Therefore, PointCLIP is a promising alternative for effective 3D point cloud understanding via CLIP under low resource cost and data regime. We conduct thorough experiments on widely-adopted ModelNet10, ModelNet40 and the challenging ScanObjectNN to demonstrate the effectiveness of PointCLIP. The code is released at https://github.com/ZrrSkywalker/PointCLIP.

Renrui Zhang, Rongyao Fang, Wei Zhang et al.

Contrastive Vision-Language Pre-training, known as CLIP, has provided a new paradigm for learning visual representations by using large-scale contrastive image-text pairs. It shows impressive performance on zero-shot knowledge transfer to downstream tasks. To further enhance CLIP's few-shot capability, CLIP-Adapter proposed to fine-tune a lightweight residual feature adapter and significantly improves the performance for few-shot classification. However, such a process still needs extra training and computational resources. In this paper, we propose \textbf{T}raining-Free CL\textbf{IP}-\textbf{Adapter} (\textbf{Tip-Adapter}), which not only inherits CLIP's training-free advantage but also performs comparably or even better than CLIP-Adapter. Tip-Adapter does not require any back propagation for training the adapter, but creates the weights by a key-value cache model constructed from the few-shot training set. In this non-parametric manner, Tip-Adapter acquires well-performed adapter weights without any training, which is both efficient and effective. Moreover, the performance of Tip-Adapter can be further boosted by fine-tuning such properly initialized adapter for only a few epochs with super-fast convergence speed. We conduct extensive experiments of few-shot classification on ImageNet and other 10 datasets to demonstrate the superiority of proposed Tip-Adapter. The code will be released at \url{https://github.com/gaopengcuhk/Tip-Adapter}.

Hang Xu, Shaoju Wang, Xinyue Cai et al.

We address the curve lane detection problem which poses more realistic challenges than conventional lane detection for better facilitating modern assisted/autonomous driving systems. Current hand-designed lane detection methods are not robust enough to capture the curve lanes especially the remote parts due to the lack of modeling both long-range contextual information and detailed curve trajectory. In this paper, we propose a novel lane-sensitive architecture search framework named CurveLane-NAS to automatically capture both long-ranged coherent and accurate short-range curve information while unifying both architecture search and post-processing on curve lane predictions via point blending. It consists of three search modules: a) a feature fusion search module to find a better fusion of the local and global context for multi-level hierarchy features; b) an elastic backbone search module to explore an efficient feature extractor with good semantics and latency; c) an adaptive point blending module to search a multi-level post-processing refinement strategy to combine multi-scale head prediction. The unified framework ensures lane-sensitive predictions by the mutual guidance between NAS and adaptive point blending. Furthermore, we also steer forward to release a more challenging benchmark named CurveLanes for addressing the most difficult curve lanes. It consists of 150K images with 680K labels.The new dataset can be downloaded at github.com/xbjxh/CurveLanes (already anonymized for this submission). Experiments on the new CurveLanes show that the SOTA lane detection methods suffer substantial performance drop while our model can still reach an 80+% F1-score. Extensive experiments on traditional lane benchmarks such as CULane also demonstrate the superiority of our CurveLane-NAS, e.g. achieving a new SOTA 74.8% F1-score on CULane.

Wei Wei, Lingjie Xu, Lingling Jin et al.

Deep neural network (DNN) architectures, such as convolutional neural networks (CNN), involve heavy computation and require hardware, such as CPU, GPU, and AI accelerators, to provide the massive computing power. With the many varieties of AI hardware prevailing on the market, it is often hard to decide which one is the best to use. Thus, benchmarking AI hardware effectively becomes important and is of great help to select and optimize AI hardware. Unfortunately, there are few AI benchmarks available in both academia and industry. Examples are BenchNN[1], DeepBench[2], and Dawn Bench[3], which are usually a collection of typical real DNN applications. While these benchmarks provide performance comparison across different AI hardware, they suffer from a number of drawbacks. First, they cannot adapt to the emerging changes of DNN algorithms and are fixed once selected. Second, they contain tens to hundreds of applications and take very long time to finish running. Third, they are mainly selected from open sources, which are restricted by copyright and are not representable to proprietary applications. In this work, a synthetic benchmarks framework is firstly proposed to address the above drawbacks of AI benchmarks. Instead of pre-selecting a set of open-sourced benchmarks and running all of them, the synthetic approach generates only a one or few benchmarks that best represent a broad range of applications using profiled workload characteristics data of these applications. Thus, it can adapt to emerging changes of new DNN algorithms by re-profiling new applications and updating itself, greatly reduce benchmark count and running time, and strongly represent DNN applications of interests. The generated benchmarks are called AI Matrix, serving as a performance benchmarks matching the statistical workload characteristics of a combination of applications of interests.

Xinpeng Ding, Jinahua Han, Hang Xu et al.

The rise of multimodal large language models (MLLMs) has spurred interest in language-based driving tasks. However, existing research typically focuses on limited tasks and often omits key multi-view and temporal information which is crucial for robust autonomous driving. To bridge these gaps, we introduce NuInstruct, a novel dataset with 91K multi-view video-QA pairs across 17 subtasks, where each task demands holistic information (e.g., temporal, multi-view, and spatial), significantly elevating the challenge level. To obtain NuInstruct, we propose a novel SQL-based method to generate instruction-response pairs automatically, which is inspired by the driving logical progression of humans. We further present BEV-InMLLM, an end-to-end method for efficiently deriving instruction-aware Bird's-Eye-View (BEV) features, language-aligned for large language models. BEV-InMLLM integrates multi-view, spatial awareness, and temporal semantics to enhance MLLMs' capabilities on NuInstruct tasks. Moreover, our proposed BEV injection module is a plug-and-play method for existing MLLMs. Our experiments on NuInstruct demonstrate that BEV-InMLLM significantly outperforms existing MLLMs, e.g. around 9% improvement on various tasks. We plan to release our NuInstruct for future research development.

Runmin Cong, Hang Xiong, Jinpeng Chen et al.

Previous Few-Shot Segmentation (FSS) approaches exclusively utilize support features for prototype generation, neglecting the specific requirements of the query. To address this, we present the Query-guided Prototype Evolution Network (QPENet), a new method that integrates query features into the generation process of foreground and background prototypes, thereby yielding customized prototypes attuned to specific queries. The evolution of the foreground prototype is accomplished through a \textit{support-query-support} iterative process involving two new modules: Pseudo-prototype Generation (PPG) and Dual Prototype Evolution (DPE). The PPG module employs support features to create an initial prototype for the preliminary segmentation of the query image, resulting in a pseudo-prototype reflecting the unique needs of the current query. Subsequently, the DPE module performs reverse segmentation on support images using this pseudo-prototype, leading to the generation of evolved prototypes, which can be considered as custom solutions. As for the background prototype, the evolution begins with a global background prototype that represents the generalized features of all training images. We also design a Global Background Cleansing (GBC) module to eliminate potential adverse components mirroring the characteristics of the current foreground class. Experimental results on the PASCAL-$5^i$ and COCO-$20^i$ datasets attest to the substantial enhancements achieved by QPENet over prevailing state-of-the-art techniques, underscoring the validity of our ideas.

Runhui Huang, Chunwei Wang, Junwei Yang et al.

We present ILLUME+ that leverages dual visual tokenization and a diffusion decoder to improve both deep semantic understanding and high-fidelity image generation. Existing unified models have struggled to simultaneously handle the three fundamental capabilities in a unified model: understanding, generation, and editing. Models like Chameleon and EMU3 utilize VQGAN for image discretization, due to the lack of deep semantic interaction, they lag behind specialist models like LLaVA in visual understanding tasks. To mitigate this, LaViT and ILLUME employ semantic encoders for tokenization, but they struggle with image editing due to poor texture preservation. Meanwhile, Janus series decouples the input and output image representation, limiting their abilities to seamlessly handle interleaved image-text understanding and generation. In contrast, ILLUME+ introduces a unified dual visual tokenizer, DualViTok, which preserves both fine-grained textures and text-aligned semantics while enabling a coarse-to-fine image representation strategy for multimodal understanding and generation. Additionally, we employ a diffusion model as the image detokenizer for enhanced generation quality and efficient super-resolution. ILLUME+ follows a continuous-input, discrete-output scheme within the unified MLLM and adopts a progressive training procedure that supports dynamic resolution across the vision tokenizer, MLLM, and diffusion decoder. This design allows for flexible and efficient context-aware image editing and generation across diverse tasks. ILLUME+ (3B) exhibits competitive performance against existing unified MLLMs and specialized models across multimodal understanding, generation, and editing benchmarks. With its strong performance, ILLUME+ provides a scalable and versatile foundation for future multimodal applications. Project Page: https://illume-unified-mllm.github.io/.

Jiacheng Zhang, Jiaming Li, Xiangru Lin et al.

We delve into pseudo-labeling for semi-supervised monocular 3D object detection (SSM3OD) and discover two primary issues: a misalignment between the prediction quality of 3D and 2D attributes and the tendency of depth supervision derived from pseudo-labels to be noisy, leading to significant optimization conflicts with other reliable forms of supervision. We introduce a novel decoupled pseudo-labeling (DPL) approach for SSM3OD. Our approach features a Decoupled Pseudo-label Generation (DPG) module, designed to efficiently generate pseudo-labels by separately processing 2D and 3D attributes. This module incorporates a unique homography-based method for identifying dependable pseudo-labels in BEV space, specifically for 3D attributes. Additionally, we present a DepthGradient Projection (DGP) module to mitigate optimization conflicts caused by noisy depth supervision of pseudo-labels, effectively decoupling the depth gradient and removing conflicting gradients. This dual decoupling strategy-at both the pseudo-label generation and gradient levels-significantly improves the utilization of pseudo-labels in SSM3OD. Our comprehensive experiments on the KITTI benchmark demonstrate the superiority of our method over existing approaches.

Yimeng Fan, Changsong Liu, Mingyang Li et al.

Spiking Neural Networks (SNNs) are the third generation of neural networks. They have gained widespread attention in object detection due to their low power consumption and biological interpretability. However, existing SNN-based object detection methods suffer from local firing saturation, where neurons in information-concentrated regions fire continuously throughout all time steps. This abnormal neuron firing pattern reduces the feature discrimination capability and detection accuracy, while also increasing the firing rates that prevent SNNs from achieving their potential energy efficiency. To address this problem, we propose SpikeDet, a novel spiking object detector that optimizes firing patterns for accurate and energy-efficient detection. Specifically, we design a spiking backbone network, MDSNet, which effectively adjusts the membrane synaptic input distribution at each layer, achieving better neuron firing patterns during spiking feature extraction. Additionally, to better utilize and preserve these high-quality backbone features, we introduce the Spiking Multi-direction Fusion Module (SMFM), which realizes multi-direction fusion of spiking features, enhancing the multi-scale detection capability of the model. Experimental results demonstrate that SpikeDet achieves superior performance. On the COCO 2017 dataset, it achieves 51.4% AP, outperforming previous SNN-based methods by 2.5% AP while requiring only half the power consumption. On object detection sub-tasks, including the GEN1 event-based dataset and the URPC 2019 underwater dataset, SpikeDet also achieves the best performance. Notably, on GEN1, our method achieves 47.6% AP, outperforming previous SNN-based methods by 7.2% AP with better energy efficiency.

Yansong Xu, Junlin Li, Wei Zhang et al.

3D gaussian splatting has advanced simultaneous localization and mapping (SLAM) technology by enabling real-time positioning and the construction of high-fidelity maps. However, the uncertainty in gaussian position and initialization parameters introduces challenges, often requiring extensive iterative convergence and resulting in redundant or insufficient gaussian representations. To address this, we introduce a novel adaptive densification method based on Fourier frequency domain analysis to establish gaussian priors for rapid convergence. Additionally, we propose constructing independent and unified sparse and dense maps, where a sparse map supports efficient tracking via Generalized Iterative Closest Point (GICP) and a dense map creates high-fidelity visual representations. This is the first SLAM system leveraging frequency domain analysis to achieve high-quality gaussian mapping in real-time. Experimental results demonstrate an average frame rate of 36 FPS on Replica and TUM RGB-D datasets, achieving competitive accuracy in both localization and mapping.

Jiaxi Gu, Xiaojun Meng, Guansong Lu et al.

Vision-Language Pre-training (VLP) models have shown remarkable performance on various downstream tasks. Their success heavily relies on the scale of pre-trained cross-modal datasets. However, the lack of large-scale datasets and benchmarks in Chinese hinders the development of Chinese VLP models and broader multilingual applications. In this work, we release a large-scale Chinese cross-modal dataset named Wukong, which contains 100 million Chinese image-text pairs collected from the web. Wukong aims to benchmark different multi-modal pre-training methods to facilitate the VLP research and community development. Furthermore, we release a group of models pre-trained with various image encoders (ViT-B/ViT-L/SwinT) and also apply advanced pre-training techniques into VLP such as locked-image text tuning, token-wise similarity in contrastive learning, and reduced-token interaction. Extensive experiments and a benchmarking of different downstream tasks including a new largest human-verified image-text test dataset are also provided. Experiments show that Wukong can serve as a promising Chinese pre-training dataset and benchmark for different cross-modal learning methods. For the zero-shot image classification task on 10 datasets, $Wukong_{ViT-L}$ achieves an average accuracy of 73.03%. For the image-text retrieval task, it achieves a mean recall of 71.6% on AIC-ICC which is 12.9% higher than WenLan 2.0. Also, our Wukong models are benchmarked on downstream tasks with other variants on multiple datasets, e.g., Flickr8K-CN, Flickr-30K-CN, COCO-CN, et al. More information can be referred to: https://wukong-dataset.github.io/wukong-dataset/.

Mohan Zhou, Yalong Bai, Wei Zhang et al.

We present a new listening head generation benchmark, for synthesizing responsive feedbacks of a listener (e.g., nod, smile) during a face-to-face conversation. As the indispensable complement to talking heads generation, listening head generation has seldomly been studied in literature. Automatically synthesizing listening behavior that actively responds to a talking head, is critical to applications such as digital human, virtual agents and social robots. In this work, we propose a novel dataset "ViCo", highlighting the listening head generation during a face-to-face conversation. A total number of 92 identities (67 speakers and 76 listeners) are involved in ViCo, featuring 483 clips in a paired "speaking-listening" pattern, where listeners show three listening styles based on their attitudes: positive, neutral, negative. Different from traditional speech-to-gesture or talking-head generation, listening head generation takes as input both the audio and visual signals from the speaker, and gives non-verbal feedbacks (e.g., head motions, facial expressions) in a real-time manner. Our dataset supports a wide range of applications such as human-to-human interaction, video-to-video translation, cross-modal understanding and generation. To encourage further research, we also release a listening head generation baseline, conditioning on different listening attitudes. Code & ViCo dataset: https://project.mhzhou.com/vico.

Wei Zhang, Mingrui Liu, Yu Feng et al.

Distributed Deep Learning (DDL) is essential for large-scale Deep Learning (DL) training. Synchronous Stochastic Gradient Descent (SSGD) 1 is the de facto DDL optimization method. Using a sufficiently large batch size is critical to achieving DDL runtime speedup. In a large batch setting, the learning rate must be increased to compensate for the reduced number of parameter updates. However, a large learning rate may harm convergence in SSGD and training could easily diverge. Recently, Decentralized Parallel SGD (DPSGD) has been proposed to improve distributed training speed. In this paper, we find that DPSGD not only has a system-wise run-time benefit but also a significant convergence benefit over SSGD in the large batch setting. Based on a detailed analysis of the DPSGD learning dynamics, we find that DPSGD introduces additional landscape-dependent noise that automatically adjusts the effective learning rate to improve convergence. In addition, we theoretically show that this noise smoothes the loss landscape, hence allowing a larger learning rate. We conduct extensive studies over 18 state-of-the-art DL models/tasks and demonstrate that DPSGD often converges in cases where SSGD diverges for large learning rates in the large batch setting. Our findings are consistent across two different application domains: Computer Vision (CIFAR10 and ImageNet-1K) and Automatic Speech Recognition (SWB300 and SWB2000), and two different types of neural network models: Convolutional Neural Networks and Long Short-Term Memory Recurrent Neural Networks.

Tong Shen, Jiawei Zuo, Fan Shi et al.

We demonstrate ViDA-MAN, a digital-human agent for multi-modal interaction, which offers realtime audio-visual responses to instant speech inquiries. Compared to traditional text or voice-based system, ViDA-MAN offers human-like interactions (e.g, vivid voice, natural facial expression and body gestures). Given a speech request, the demonstration is able to response with high quality videos in sub-second latency. To deliver immersive user experience, ViDA-MAN seamlessly integrates multi-modal techniques including Acoustic Speech Recognition (ASR), multi-turn dialog, Text To Speech (TTS), talking heads video generation. Backed with large knowledge base, ViDA-MAN is able to chat with users on a number of topics including chit-chat, weather, device control, News recommendations, booking hotels, as well as answering questions via structured knowledge.

Jie Wu, Wei Zhang, Guanbin Li et al.

In this paper, we introduce a novel task, referred to as Weakly-Supervised Spatio-Temporal Anomaly Detection (WSSTAD) in surveillance video. Specifically, given an untrimmed video, WSSTAD aims to localize a spatio-temporal tube (i.e., a sequence of bounding boxes at consecutive times) that encloses the abnormal event, with only coarse video-level annotations as supervision during training. To address this challenging task, we propose a dual-branch network which takes as input the proposals with multi-granularities in both spatial-temporal domains. Each branch employs a relationship reasoning module to capture the correlation between tubes/videolets, which can provide rich contextual information and complex entity relationships for the concept learning of abnormal behaviors. Mutually-guided Progressive Refinement framework is set up to employ dual-path mutual guidance in a recurrent manner, iteratively sharing auxiliary supervision information across branches. It impels the learned concepts of each branch to serve as a guide for its counterpart, which progressively refines the corresponding branch and the whole framework. Furthermore, we contribute two datasets, i.e., ST-UCF-Crime and STRA, consisting of videos containing spatio-temporal abnormal annotations to serve as the benchmarks for WSSTAD. We conduct extensive qualitative and quantitative evaluations to demonstrate the effectiveness of the proposed approach and analyze the key factors that contribute more to handle this task.

Yalong Bai, Mohan Zhou, Wei Zhang et al.

Data augmentation is practically helpful for visual recognition, especially at the time of data scarcity. However, such success is only limited to quite a few light augmentations (e.g., random crop, flip). Heavy augmentations are either unstable or show adverse effects during training, owing to the big gap between the original and augmented images. This paper introduces a novel network design, noted as Augmentation Pathways (AP), to systematically stabilize training on a much wider range of augmentation policies. Notably, AP tames various heavy data augmentations and stably boosts performance without a careful selection among augmentation policies. Unlike traditional single pathway, augmented images are processed in different neural paths. The main pathway handles the light augmentations, while other pathways focus on the heavier augmentations. By interacting with multiple paths in a dependent manner, the backbone network robustly learns from shared visual patterns among augmentations, and suppresses the side effect of heavy augmentations at the same time. Furthermore, we extend AP to high-order versions for high-order scenarios, demonstrating its robustness and flexibility in practical usage. Experimental results on ImageNet demonstrate the compatibility and effectiveness on a much wider range of augmentations, while consuming fewer parameters and lower computational costs at inference time.

Tianyu Hua, Hongdong Zheng, Yalong Bai et al.

The significant progress on Generative Adversarial Networks (GANs) has facilitated realistic single-object image generation based on language input. However, complex-scene generation (with various interactions among multiple objects) still suffers from messy layouts and object distortions, due to diverse configurations in layouts and appearances. Prior methods are mostly object-driven and ignore their inter-relations that play a significant role in complex-scene images. This work explores relationship-aware complex-scene image generation, where multiple objects are inter-related as a scene graph. With the help of relationships, we propose three major updates in the generation framework. First, reasonable spatial layouts are inferred by jointly considering the semantics and relationships among objects. Compared to standard location regression, we show relative scales and distances serve a more reliable target. Second, since the relations between objects significantly influence an object's appearance, we design a relation-guided generator to generate objects reflecting their relationships. Third, a novel scene graph discriminator is proposed to guarantee the consistency between the generated image and the input scene graph. Our method tends to synthesize plausible layouts and objects, respecting the interplay of multiple objects in an image. Experimental results on Visual Genome and HICO-DET datasets show that our proposed method significantly outperforms prior arts in terms of IS and FID metrics. Based on our user study and visual inspection, our method is more effective in generating logical layout and appearance for complex-scenes.

Xuanyu He, Wei Zhang, Ran Song et al.

Unsupervised person re-identification (re-ID) aims at closing the performance gap to supervised methods. These methods build reliable relationship between data points while learning representations. However, we empirically show that the reason why they are successful is not only their label generation mechanisms, but also their unexplored designs. By studying two unsupervised person re-ID methods in a cross-method way, we point out a hard negative problem is handled implicitly by their designs of data augmentations and PK sampler respectively. In this paper, we find another simple solution for the problem, i.e., taking more positives during training, by which we generate pseudo-labels and update models in an iterative manner. Based on our findings, we propose a contrastive learning method without a memory back for unsupervised person re-ID. Our method works well on benchmark datasets and outperforms the state-of-the-art methods. Code will be made available.

Qiuyu Chen, Wei Zhang, Jianping Fan

Instance-level alignment is widely exploited for person re-identification, e.g. spatial alignment, latent semantic alignment and triplet alignment. This paper probes another feature alignment modality, namely cluster-level feature alignment across whole dataset, where the model can see not only the sampled images in local mini-batch but the global feature distribution of the whole dataset from distilled anchors. Towards this aim, we propose anchor loss and investigate many variants of cluster-level feature alignment, which consists of iterative aggregation and alignment from the overview of dataset. Our extensive experiments have demonstrated that our methods can provide consistent and significant performance improvement with small training efforts after the saturation of traditional training. In both theoretical and experimental aspects, our proposed methods can result in more stable and guided optimization towards better representation and generalization for well-aligned embedding.

Weidong Zhang, Wei Zhang, Yinda Zhang

The task of room layout estimation is to locate the wall-floor, wall-ceiling, and wall-wall boundaries. Most recent methods solve this problem based on edge/keypoint detection or semantic segmentation. However, these approaches have shown limited attention on the geometry of the dominant planes and the intersection between them, which has significant impact on room layout. In this work, we propose to incorporate geometric reasoning to deep learning for layout estimation. Our approach learns to infer the depth maps of the dominant planes in the scene by predicting the pixel-level surface parameters, and the layout can be generated by the intersection of the depth maps. Moreover, we present a new dataset with pixel-level depth annotation of dominant planes. It is larger than the existing datasets and contains both cuboid and non-cuboid rooms. Experimental results show that our approach produces considerable performance gains on both 2D and 3D datasets.

Tao Mei, Wei Zhang, Ting Yao

Vision and language are two fundamental capabilities of human intelligence. Humans routinely perform tasks through the interactions between vision and language, supporting the uniquely human capacity to talk about what they see or hallucinate a picture on a natural-language description. The valid question of how language interacts with vision motivates us researchers to expand the horizons of computer vision area. In particular, "vision to language" is probably one of the most popular topics in the past five years, with a significant growth in both volume of publications and extensive applications, e.g., captioning, visual question answering, visual dialog, language navigation, etc. Such tasks boost visual perception with more comprehensive understanding and diverse linguistic representations. Going beyond the progresses made in "vision to language," language can also contribute to vision understanding and offer new possibilities of visual content creation, i.e., "language to vision." The process performs as a prism through which to create visual content conditioning on the language inputs. This paper reviews the recent advances along these two dimensions: "vision to language" and "language to vision." More concretely, the former mainly focuses on the development of image/video captioning, as well as typical encoder-decoder structures and benchmarks, while the latter summarizes the technologies of visual content creation. The real-world deployment or services of vision and language are elaborated as well.

Shuo Yang, Wei Zhang, Weizhi Lu et al.

Learning actions from human demonstration is an emerging trend for designing intelligent robotic systems, which can be referred as video to command. The performance of such approach highly relies on the quality of video captioning. However, the general video captioning methods focus more on the understanding of the full frame, lacking of consideration on the specific object of interests in robotic manipulations. We propose a novel deep model to learn actions from human demonstration video for robotic manipulation. It consists of two deep networks, grasp detection network (GNet) and video captioning network (CNet). GNet performs two functions: providing grasp solutions and extracting the local features for the object of interests in robotic manipulation. CNet outputs the captioning results by fusing the features of both full frames and local objects. Experimental results on UR5 robotic arm show that our method could produce more accurate command from video demonstration than state-of-the-art work, thereby leading to more robust grasping performance.

Tong Shen, Dong Gong, Wei Zhang et al.

Training a semantic segmentation model requires a large amount of pixel-level annotation, hampering its application at scale. With computer graphics, we can generate almost unlimited training data with precise annotation. However,a deep model trained with synthetic data usually cannot directly generalize well to realistic images due to domain shift. It has been observed that highly confident labels for the unlabeled real images may be predicted relying on the labeled synthetic data. To tackle the unsupervised domain adaptation problem, we explore the possibilities to generate high-quality labels as proxy labels to supervise the training on target data. Specifically, we propose a novel proxy-based method using multi-adversarial training. We first train the model using synthetic data (source domain). Multiple discriminators are used to align the features be-tween the source and target domain (real images) at different levels. Then we focus on obtaining and selecting high-quality proxy labels by incorporating both the confidence of the class predictor and that from the adversarial discriminators. Our discriminators not only work as a regularizer to encourage feature alignment but also provide an alternative confidence measure for generating proxy labels. Relying on the generated high-quality proxies, our model can be trained in a "supervised manner" on the target do-main. On two major tasks, GTA5->Cityscapes and SYNTHIA->Cityscapes, our method achieves state-of-the-art results, outperforming the previous by a large margin.

Danlu Chen, Xu-Yao Zhang, Wei Zhang et al.

Deep learning based approaches have achieved significant progresses in different tasks like classification, detection, segmentation, and so on. Ensemble learning is widely known to further improve performance by combining multiple complementary models. It is easy to apply ensemble learning for classification tasks, for example, based on averaging, voting, or other methods. However, for other tasks (like object detection) where the outputs are varying in quantity and unable to be simply compared, the ensemble of multiple models become difficult. In this paper, we propose a new method called Predictive Ensemble Learning (PEL), based on powerful predictive ability of deep neural networks, to directly predict the best performing model among a pool of base models for each test example, thus transforming ensemble learning to a traditional classification task. Taking scene text detection as the application, where no suitable ensemble learning strategy exists, PEL can significantly improve the performance, compared to either individual state-of-the-art models, or the fusion of multiple models by non-maximum suppression. Experimental results show the possibility and potential of PEL in predicting different models' performance based only on a query example, which can be extended for ensemble learning in many other complex tasks.

Xinyue Zhang, Yanfang Wang, Wei Zhang et al.

Cosmological surveys aim at answering fundamental questions about our Universe, including the nature of dark matter or the reason of unexpected accelerated expansion of the Universe. In order to answer these questions, two important ingredients are needed: 1) data from observations and 2) a theoretical model that allows fast comparison between observation and theory. Most of the cosmological surveys observe galaxies, which are very difficult to model theoretically due to the complicated physics involved in their formation and evolution; modeling realistic galaxies over cosmological volumes requires running computationally expensive hydrodynamic simulations that can cost millions of CPU hours. In this paper, we propose to use deep learning to establish a mapping between the 3D galaxy distribution in hydrodynamic simulations and its underlying dark matter distribution. One of the major challenges in this pursuit is the very high sparsity in the predicted galaxy distribution. To this end, we develop a two-phase convolutional neural network architecture to generate fast galaxy catalogues, and compare our results against a standard cosmological technique. We find that our proposed approach either outperforms or is competitive with traditional cosmological techniques. Compared to the common methods used in cosmology, our approach also provides a nice trade-off between time-consumption (comparable to fastest benchmark in the literature) and the quality and accuracy of the predicted simulation. In combination with current and upcoming data from cosmological observations, our method has the potential to answer fundamental questions about our Universe with the highest accuracy.

Chuhui Xue, Shijian Lu, Wei Zhang

State-of-the-art scene text detection techniques predict quadrilateral boxes that are prone to localization errors while dealing with straight or curved text lines of different orientations and lengths in scenes. This paper presents a novel multi-scale shape regression network (MSR) that is capable of locating text lines of different lengths, shapes and curvatures in scenes. The proposed MSR detects scene texts by predicting dense text boundary points that inherently capture the location and shape of text lines accurately and are also more tolerant to the variation of text line length as compared with the state of the arts using proposals or segmentation. Additionally, the multi-scale network extracts and fuses features at different scales which demonstrates superb tolerance to the text scale variation. Extensive experiments over several public datasets show that the proposed MSR obtains superior detection performance for both curved and straight text lines of different lengths and orientations.

Xiangru Lian, Ce Zhang, Huan Zhang et al.

Most distributed machine learning systems nowadays, including TensorFlow and CNTK, are built in a centralized fashion. One bottleneck of centralized algorithms lies on high communication cost on the central node. Motivated by this, we ask, can decentralized algorithms be faster than its centralized counterpart? Although decentralized PSGD (D-PSGD) algorithms have been studied by the control community, existing analysis and theory do not show any advantage over centralized PSGD (C-PSGD) algorithms, simply assuming the application scenario where only the decentralized network is available. In this paper, we study a D-PSGD algorithm and provide the first theoretical analysis that indicates a regime in which decentralized algorithms might outperform centralized algorithms for distributed stochastic gradient descent. This is because D-PSGD has comparable total computational complexities to C-PSGD but requires much less communication cost on the busiest node. We further conduct an empirical study to validate our theoretical analysis across multiple frameworks (CNTK and Torch), different network configurations, and computation platforms up to 112 GPUs. On network configurations with low bandwidth or high latency, D-PSGD can be up to one order of magnitude faster than its well-optimized centralized counterparts.

Wei Zhang, Shengnan Hu, Kan Liu et al.

This paper presents a novel approach for video-based person re-identification using multiple Convolutional Neural Networks (CNNs). Unlike previous work, we intend to extract a compact yet discriminative appearance representation from several frames rather than the whole sequence. Specifically, given a video, the representative frames are selected based on the walking profile of consecutive frames. A multiple CNN architecture incorporated with feature pooling is proposed to learn and compile the features of the selected representative frames into a compact description about the pedestrian for identification. Experiments are conducted on benchmark datasets to demonstrate the superiority of the proposed method over existing person re-identification approaches.