Hao Tang

Guolei Sun, Yun Liu, Hao Tang et al.

The essence of video semantic segmentation (VSS) is how to leverage temporal information for prediction. Previous efforts are mainly devoted to developing new techniques to calculate the cross-frame affinities such as optical flow and attention. Instead, this paper contributes from a different angle by mining relations among cross-frame affinities, upon which better temporal information aggregation could be achieved. We explore relations among affinities in two aspects: single-scale intrinsic correlations and multi-scale relations. Inspired by traditional feature processing, we propose Single-scale Affinity Refinement (SAR) and Multi-scale Affinity Aggregation (MAA). To make it feasible to execute MAA, we propose a Selective Token Masking (STM) strategy to select a subset of consistent reference tokens for different scales when calculating affinities, which also improves the efficiency of our method. At last, the cross-frame affinities strengthened by SAR and MAA are adopted for adaptively aggregating temporal information. Our experiments demonstrate that the proposed method performs favorably against state-of-the-art VSS methods. The code is publicly available at https://github.com/GuoleiSun/VSS-MRCFA

Kai Zhang, Yawei Li, Jingyun Liang et al. · eth-zurich

While recent years have witnessed a dramatic upsurge of exploiting deep neural networks toward solving image denoising, existing methods mostly rely on simple noise assumptions, such as additive white Gaussian noise (AWGN), JPEG compression noise and camera sensor noise, and a general-purpose blind denoising method for real images remains unsolved. In this paper, we attempt to solve this problem from the perspective of network architecture design and training data synthesis. Specifically, for the network architecture design, we propose a swin-conv block to incorporate the local modeling ability of residual convolutional layer and non-local modeling ability of swin transformer block, and then plug it as the main building block into the widely-used image-to-image translation UNet architecture. For the training data synthesis, we design a practical noise degradation model which takes into consideration different kinds of noise (including Gaussian, Poisson, speckle, JPEG compression, and processed camera sensor noises) and resizing, and also involves a random shuffle strategy and a double degradation strategy. Extensive experiments on AGWN removal and real image denoising demonstrate that the new network architecture design achieves state-of-the-art performance and the new degradation model can help to significantly improve the practicability. We believe our work can provide useful insights into current denoising research.

Neng Dong, Liyan Zhang, Shuanglin Yan et al.

Occlusion perturbation presents a significant challenge in person re-identification (re-ID), and existing methods that rely on external visual cues require additional computational resources and only consider the issue of missing information caused by occlusion. In this paper, we propose a simple yet effective framework, termed Erasing, Transforming, and Noising Defense Network (ETNDNet), which treats occlusion as a noise disturbance and solves occluded person re-ID from the perspective of adversarial defense. In the proposed ETNDNet, we introduce three strategies: Firstly, we randomly erase the feature map to create an adversarial representation with incomplete information, enabling adversarial learning of identity loss to protect the re-ID system from the disturbance of missing information. Secondly, we introduce random transformations to simulate the position misalignment caused by occlusion, training the extractor and classifier adversarially to learn robust representations immune to misaligned information. Thirdly, we perturb the feature map with random values to address noisy information introduced by obstacles and non-target pedestrians, and employ adversarial gaming in the re-ID system to enhance its resistance to occlusion noise. Without bells and whistles, ETNDNet has three key highlights: (i) it does not require any external modules with parameters, (ii) it effectively handles various issues caused by occlusion from obstacles and non-target pedestrians, and (iii) it designs the first GAN-based adversarial defense paradigm for occluded person re-ID. Extensive experiments on five public datasets fully demonstrate the effectiveness, superiority, and practicality of the proposed ETNDNet. The code will be released at \url{https://github.com/nengdong96/ETNDNet}.

Wenhao Li, Mengyuan Liu, Hong Liu et al.

Modern multi-layer perceptron (MLP) models have shown competitive results in learning visual representations without self-attention. However, existing MLP models are not good at capturing local details and lack prior knowledge of human body configurations, which limits their modeling power for skeletal representation learning. To address these issues, we propose a simple yet effective graph-reinforced MLP-Like architecture, named GraphMLP, that combines MLPs and graph convolutional networks (GCNs) in a global-local-graphical unified architecture for 3D human pose estimation. GraphMLP incorporates the graph structure of human bodies into an MLP model to meet the domain-specific demand of the 3D human pose, while allowing for both local and global spatial interactions. Furthermore, we propose to flexibly and efficiently extend the GraphMLP to the video domain and show that complex temporal dynamics can be effectively modeled in a simple way with negligible computational cost gains in the sequence length. To the best of our knowledge, this is the first MLP-Like architecture for 3D human pose estimation in a single frame and a video sequence. Extensive experiments show that the proposed GraphMLP achieves state-of-the-art performance on two datasets, i.e., Human3.6M and MPI-INF-3DHP. Code and models are available at https://github.com/Vegetebird/GraphMLP.

Shuanglin Yan, Hao Tang, Liyan Zhang et al.

Text-based person search (TBPS) is a challenging task that aims to search pedestrian images with the same identity from an image gallery given a query text. In recent years, TBPS has made remarkable progress and state-of-the-art methods achieve superior performance by learning local fine-grained correspondence between images and texts. However, most existing methods rely on explicitly generated local parts to model fine-grained correspondence between modalities, which is unreliable due to the lack of contextual information or the potential introduction of noise. Moreover, existing methods seldom consider the information inequality problem between modalities caused by image-specific information. To address these limitations, we propose an efficient joint Multi-level Alignment Network (MANet) for TBPS, which can learn aligned image/text feature representations between modalities at multiple levels, and realize fast and effective person search. Specifically, we first design an image-specific information suppression module, which suppresses image background and environmental factors by relation-guided localization and channel attention filtration respectively. This module effectively alleviates the information inequality problem and realizes the alignment of information volume between images and texts. Secondly, we propose an implicit local alignment module to adaptively aggregate all pixel/word features of image/text to a set of modality-shared semantic topic centers and implicitly learn the local fine-grained correspondence between modalities without additional supervision and cross-modal interactions. And a global alignment is introduced as a supplement to the local perspective. The cooperation of global and local alignment modules enables better semantic alignment between modalities. Extensive experiments on multiple databases demonstrate the effectiveness and superiority of our MANet.

Neng Dong, Shuanglin Yan, Hao Tang et al.

Occluded person re-identification (re-ID) presents a challenging task due to occlusion perturbations. Although great efforts have been made to prevent the model from being disturbed by occlusion noise, most current solutions only capture information from a single image, disregarding the rich complementary information available in multiple images depicting the same pedestrian. In this paper, we propose a novel framework called Multi-view Information Integration and Propagation (MVI$^{2}$P). Specifically, realizing the potential of multi-view images in effectively characterizing the occluded target pedestrian, we integrate feature maps of which to create a comprehensive representation. During this process, to avoid introducing occlusion noise, we develop a CAMs-aware Localization module that selectively integrates information contributing to the identification. Additionally, considering the divergence in the discriminative nature of different images, we design a probability-aware Quantification module to emphatically integrate highly reliable information. Moreover, as multiple images with the same identity are not accessible in the testing stage, we devise an Information Propagation (IP) mechanism to distill knowledge from the comprehensive representation to that of a single occluded image. Extensive experiments and analyses have unequivocally demonstrated the effectiveness and superiority of the proposed MVI$^{2}$P. The code will be released at \url{https://github.com/nengdong96/MVIIP}.

Hao Tang, Kevin Liang, Matt Feiszli et al.

Visual object tracking is a key component to many egocentric vision problems. However, the full spectrum of challenges of egocentric tracking faced by an embodied AI is underrepresented in many existing datasets; these tend to focus on relatively short, third-person videos. Egocentric video has several distinguishing characteristics from those commonly found in past datasets: frequent large camera motions and hand interactions with objects commonly lead to occlusions or objects exiting the frame, and object appearance can change rapidly due to widely different points of view, scale, or object states. Embodied tracking is also naturally long-term, and being able to consistently (re-)associate objects to their appearances and disappearances over as long as a lifetime is critical. Previous datasets under-emphasize this re-detection problem, and their "framed" nature has led to adoption of various spatiotemporal priors that we find do not necessarily generalize to egocentric video. We thus introduce EgoTracks, a new dataset for long-term egocentric visual object tracking. Sourced from the Ego4D dataset, this new dataset presents a significant challenge to recent state-of-the-art single-object tracking models, which we find score poorly on traditional tracking metrics for our new dataset, compared to popular benchmarks. We further show improvements that can be made to a STARK tracker to significantly increase its performance on egocentric data, resulting in a baseline model we call EgoSTARK. We publicly release our annotations and benchmark, hoping our dataset leads to further advancements in tracking.

Hao Tang, Lei Ding, Songsong Wu et al.

Video processing and analysis have become an urgent task since a huge amount of videos (e.g., Youtube, Hulu) are uploaded online every day. The extraction of representative key frames from videos is very important in video processing and analysis since it greatly reduces computing resources and time. Although great progress has been made recently, large-scale video classification remains an open problem, as the existing methods have not well balanced the performance and efficiency simultaneously. To tackle this problem, this work presents an unsupervised method to retrieve the key frames, which combines Convolutional Neural Network (CNN) and Temporal Segment Density Peaks Clustering (TSDPC). The proposed TSDPC is a generic and powerful framework and it has two advantages compared with previous works, one is that it can calculate the number of key frames automatically. The other is that it can preserve the temporal information of the video. Thus it improves the efficiency of video classification. Furthermore, a Long Short-Term Memory network (LSTM) is added on the top of the CNN to further elevate the performance of classification. Moreover, a weight fusion strategy of different input networks is presented to boost the performance. By optimizing both video classification and key frame extraction simultaneously, we achieve better classification performance and higher efficiency. We evaluate our method on two popular datasets (i.e., HMDB51 and UCF101) and the experimental results consistently demonstrate that our strategy achieves competitive performance and efficiency compared with the state-of-the-art approaches.

Daqian Shi, Xiaolei Diao, Hao Tang et al.

Constructing high-quality character image datasets is challenging because real-world images are often affected by image degradation. There are limitations when applying current image restoration methods to such real-world character images, since (i) the categories of noise in character images are different from those in general images; (ii) real-world character images usually contain more complex image degradation, e.g., mixed noise at different noise levels. To address these problems, we propose a real-world character restoration network (RCRN) to effectively restore degraded character images, where character skeleton information and scale-ensemble feature extraction are utilized to obtain better restoration performance. The proposed method consists of a skeleton extractor (SENet) and a character image restorer (CiRNet). SENet aims to preserve the structural consistency of the character and normalize complex noise. Then, CiRNet reconstructs clean images from degraded character images and their skeletons. Due to the lack of benchmarks for real-world character image restoration, we constructed a dataset containing 1,606 character images with real-world degradation to evaluate the validity of the proposed method. The experimental results demonstrate that RCRN outperforms state-of-the-art methods quantitatively and qualitatively.

Xuan Shen, Yaohua Wang, Ming Lin et al.

The rapid advances in Vision Transformer (ViT) refresh the state-of-the-art performances in various vision tasks, overshadowing the conventional CNN-based models. This ignites a few recent striking-back research in the CNN world showing that pure CNN models can achieve as good performance as ViT models when carefully tuned. While encouraging, designing such high-performance CNN models is challenging, requiring non-trivial prior knowledge of network design. To this end, a novel framework termed Mathematical Architecture Design for Deep CNN (DeepMAD) is proposed to design high-performance CNN models in a principled way. In DeepMAD, a CNN network is modeled as an information processing system whose expressiveness and effectiveness can be analytically formulated by their structural parameters. Then a constrained mathematical programming (MP) problem is proposed to optimize these structural parameters. The MP problem can be easily solved by off-the-shelf MP solvers on CPUs with a small memory footprint. In addition, DeepMAD is a pure mathematical framework: no GPU or training data is required during network design. The superiority of DeepMAD is validated on multiple large-scale computer vision benchmark datasets. Notably on ImageNet-1k, only using conventional convolutional layers, DeepMAD achieves 0.7% and 1.5% higher top-1 accuracy than ConvNeXt and Swin on Tiny level, and 0.8% and 0.9% higher on Small level.

Anh Thai, Weiyao Wang, Hao Tang et al.

3D object part segmentation is essential in computer vision applications. While substantial progress has been made in 2D object part segmentation, the 3D counterpart has received less attention, in part due to the scarcity of annotated 3D datasets, which are expensive to collect. In this work, we propose to leverage a few annotated 3D shapes or richly annotated 2D datasets to perform 3D object part segmentation. We present our novel approach, termed 3-By-2 that achieves SOTA performance on different benchmarks with various granularity levels. By using features from pretrained foundation models and exploiting semantic and geometric correspondences, we are able to overcome the challenges of limited 3D annotations. Our approach leverages available 2D labels, enabling effective 3D object part segmentation. Our method 3-By-2 can accommodate various part taxonomies and granularities, demonstrating interesting part label transfer ability across different object categories. Project website: \url{https://ngailapdi.github.io/projects/3by2/}.

Pan Xie, Qipeng Zhang, Taiyi Peng et al.

The Sign Language Production (SLP) project aims to automatically translate spoken languages into sign sequences. Our approach focuses on the transformation of sign gloss sequences into their corresponding sign pose sequences (G2P). In this paper, we present a novel solution for this task by converting the continuous pose space generation problem into a discrete sequence generation problem. We introduce the Pose-VQVAE framework, which combines Variational Autoencoders (VAEs) with vector quantization to produce a discrete latent representation for continuous pose sequences. Additionally, we propose the G2P-DDM model, a discrete denoising diffusion architecture for length-varied discrete sequence data, to model the latent prior. To further enhance the quality of pose sequence generation in the discrete space, we present the CodeUnet model to leverage spatial-temporal information. Lastly, we develop a heuristic sequential clustering method to predict variable lengths of pose sequences for corresponding gloss sequences. Our results show that our model outperforms state-of-the-art G2P models on the public SLP evaluation benchmark. For more generated results, please visit our project page: \textcolor{blue}{\url{https://slpdiffusier.github.io/g2p-ddm}}

Xuan Wang, Hao Tang, Zhigang Zhu

Various contextual information has been employed by many approaches for visual detection tasks. However, most of the existing approaches only focus on specific context for specific tasks. In this paper, GMC, a general framework is proposed for multistage context learning and utilization, with various deep network architectures for various visual detection tasks. The GMC framework encompasses three stages: preprocessing, training, and post-processing. In the preprocessing stage, the representation of local context is enhanced by utilizing commonly used labeling standards. During the training stage, semantic context information is fused with visual information, leveraging prior knowledge from the training dataset to capture semantic relationships. In the post-processing stage, general topological relations and semantic masks for stuff are incorporated to enable spatial context reasoning between objects. The proposed framework provides a comprehensive and adaptable solution for context learning and utilization in visual detection scenarios. The framework offers flexibility with user-defined configurations and provide adaptability to diverse network architectures and visual detection tasks, offering an automated and streamlined solution that minimizes user effort and inference time in context learning and reasoning. Experimental results on the visual detection tasks, for storefront object detection, pedestrian detection and COCO object detection, demonstrate that our framework outperforms previous state-of-the-art detectors and transformer architectures. The experiments also demonstrate that three contextual learning components can not only be applied individually and in combination, but can also be applied to various network architectures, and its flexibility and effectiveness in various detection scenarios.

Hao Tang, Chenwei Xie, Xiaoyi Bao et al.

In this paper, we propose UniLIP, a unified framework that adapts CLIP for multimodal understanding, generation and editing. Although CLIP excels at understanding, it lacks reconstruction abilities required to be a unified visual encoder. However, previous CLIP-based unified methods fail to balance understanding and reconstruction, leading to semantic degradation or inconsistent reconstructions. In contrast, we introduce a novel two-stage training scheme with a self-distillation strategy that progressively endows CLIP with high-fidelity reconstruction abilities while preserving its original comprehension performance. For enhanced reasoning and consistency in generation and editing, we further develop a dual-condition architecture built upon the MetaQuery framework. Our architecture jointly utilizes multimodal hidden states for rich contextual details and learnable query embeddings to harness the powerful reasoning abilities of Multimodal Large Language Models (MLLMs). Leveraging advanced image representation and architectural design, UniLIP demonstrates superior instruction following and edit fidelity. With only 1B and 3B parameters, UniLIP can outperform larger unified models such as BAGEL (7B) and Uniworld-V1 (12B), achieving state-of-the-art performance of 0.90 on GenEval, 0.63 on WISE, and 3.94 on ImgEdit. These results demonstrate that UniLIP successfully expands the application of CLIP, establishing its continuous features to not only serve as the optimal choice for understanding tasks but also achieve highly competitive performance in generation and editing tasks. Code and models are available at https://github.com/nnnth/UniLIP.

Jianing Yang, Alexander Sax, Kevin J. Liang et al.

Multi-view 3D reconstruction remains a core challenge in computer vision, particularly in applications requiring accurate and scalable representations across diverse perspectives. Current leading methods such as DUSt3R employ a fundamentally pairwise approach, processing images in pairs and necessitating costly global alignment procedures to reconstruct from multiple views. In this work, we propose Fast 3D Reconstruction (Fast3R), a novel multi-view generalization to DUSt3R that achieves efficient and scalable 3D reconstruction by processing many views in parallel. Fast3R's Transformer-based architecture forwards N images in a single forward pass, bypassing the need for iterative alignment. Through extensive experiments on camera pose estimation and 3D reconstruction, Fast3R demonstrates state-of-the-art performance, with significant improvements in inference speed and reduced error accumulation. These results establish Fast3R as a robust alternative for multi-view applications, offering enhanced scalability without compromising reconstruction accuracy.

Tong Xu, Micol Spitale, Hao Tang et al.

Generating facial reactions in a human-human dyadic interaction is complex and highly dependent on the context since more than one facial reactions can be appropriate for the speaker's behaviour. This has challenged existing machine learning (ML) methods, whose training strategies enforce models to reproduce a specific (not multiple) facial reaction from each input speaker behaviour. This paper proposes the first multiple appropriate facial reaction generation framework that re-formulates the one-to-many mapping facial reaction generation problem as a one-to-one mapping problem. This means that we approach this problem by considering the generation of a distribution of the listener's appropriate facial reactions instead of multiple different appropriate facial reactions, i.e., 'many' appropriate facial reaction labels are summarised as 'one' distribution label during training. Our model consists of a perceptual processor, a cognitive processor, and a motor processor. The motor processor is implemented with a novel Reversible Multi-dimensional Edge Graph Neural Network (REGNN). This allows us to obtain a distribution of appropriate real facial reactions during the training process, enabling the cognitive processor to be trained to predict the appropriate facial reaction distribution. At the inference stage, the REGNN decodes an appropriate facial reaction by using this distribution as input. Experimental results demonstrate that our approach outperforms existing models in generating more appropriate, realistic, and synchronized facial reactions. The improved performance is largely attributed to the proposed appropriate facial reaction distribution learning strategy and the use of a REGNN. The code is available at https://github.com/TongXu-05/REGNN-Multiple-Appropriate-Facial-Reaction-Generation.

Dejia Xu, Yifan Jiang, Chen Huang et al.

In recent years there have been remarkable breakthroughs in image-to-video generation. However, the 3D consistency and camera controllability of generated frames have remained unsolved. Recent studies have attempted to incorporate camera control into the generation process, but their results are often limited to simple trajectories or lack the ability to generate consistent videos from multiple distinct camera paths for the same scene. To address these limitations, we introduce Cavia, a novel framework for camera-controllable, multi-view video generation, capable of converting an input image into multiple spatiotemporally consistent videos. Our framework extends the spatial and temporal attention modules into view-integrated attention modules, improving both viewpoint and temporal consistency. This flexible design allows for joint training with diverse curated data sources, including scene-level static videos, object-level synthetic multi-view dynamic videos, and real-world monocular dynamic videos. To our best knowledge, Cavia is the first of its kind that allows the user to precisely specify camera motion while obtaining object motion. Extensive experiments demonstrate that Cavia surpasses state-of-the-art methods in terms of geometric consistency and perceptual quality. Project Page: https://ir1d.github.io/Cavia/

Xu Ma, Peize Sun, Haoyu Ma et al.

Autoregressive (AR) models, long dominant in language generation, are increasingly applied to image synthesis but are often considered less competitive than Diffusion-based models. A primary limitation is the substantial number of image tokens required for AR models, which constrains both training and inference efficiency, as well as image resolution. To address this, we present Token-Shuffle, a novel yet simple method that reduces the number of image tokens in Transformer. Our key insight is the dimensional redundancy of visual vocabularies in Multimodal Large Language Models (MLLMs), where low-dimensional visual codes from visual encoder are directly mapped to high-dimensional language vocabularies. Leveraging this, we consider two key operations: token-shuffle, which merges spatially local tokens along channel dimension to decrease the input token number, and token-unshuffle, which untangles the inferred tokens after Transformer blocks to restore the spatial arrangement for output. Jointly training with textual prompts, our strategy requires no additional pretrained text-encoder and enables MLLMs to support extremely high-resolution image synthesis in a unified next-token prediction way while maintaining efficient training and inference. For the first time, we push the boundary of AR text-to-image generation to a resolution of 2048x2048 with gratifying generation performance. In GenAI-benchmark, our 2.7B model achieves 0.77 overall score on hard prompts, outperforming AR models LlamaGen by 0.18 and diffusion models LDM by 0.15. Exhaustive large-scale human evaluations also demonstrate our prominent image generation ability in terms of text-alignment, visual flaw, and visual appearance. We hope that Token-Shuffle can serve as a foundational design for efficient high-resolution image generation within MLLMs.

Rohit Saxena, Hao Tang, Frank Keller

Training transformer-based encoder-decoder models for long document summarization poses a significant challenge due to the quadratic memory consumption during training. Several approaches have been proposed to extend the input length at test time, but training with these approaches is still difficult, requiring truncation of input documents and causing a mismatch between training and test conditions. In this work, we propose CachED (Gradient $\textbf{Cach}$ing for $\textbf{E}$ncoder-$\textbf{D}$ecoder models), an approach that enables end-to-end training of existing transformer-based encoder-decoder models, using the entire document without truncation. Specifically, we apply non-overlapping sliding windows to input documents, followed by fusion in decoder. During backpropagation, the gradients are cached at the decoder and are passed through the encoder in chunks by re-computing the hidden vectors, similar to gradient checkpointing. In the experiments on long document summarization, we extend BART to CachED BART, processing more than 500K tokens during training and achieving superior performance without using any additional parameters.

Bin Xie, Hao Tang, Bin Duan et al.

Segment Anything Model (SAM), a prompt-driven foundation model for natural image segmentation, has demonstrated impressive zero-shot performance. However, SAM does not work when directly applied to medical image segmentation, since SAM lacks the ability to predict semantic labels, requires additional prompts, and presents suboptimal performance. Following the above issues, we propose MaskSAM, a novel mask classification prompt-free SAM adaptation framework for medical image segmentation. We design a prompt generator combined with the image encoder in SAM to generate a set of auxiliary classifier tokens, auxiliary binary masks, and auxiliary bounding boxes. Each pair of auxiliary mask and box prompts can solve the requirements of extra prompts. The semantic label prediction can be addressed by the sum of the auxiliary classifier tokens and the learnable global classifier tokens in the mask decoder of SAM. Meanwhile, we design a 3D depth-convolution adapter for image embeddings and a 3D depth-MLP adapter for prompt embeddings to efficiently fine-tune SAM. Our method achieves state-of-the-art performance on AMOS2022, 90.52% Dice, which improved by 2.7% compared to nnUNet. Our method surpasses nnUNet by 1.7% on ACDC and 1.0% on Synapse datasets.

Guohao Huo, Ruiting Dai, Jinliang Liu et al.

In deep-sea exploration and surgical robotics scenarios, environmental lighting and device resolution limitations often cause high-frequency feature attenuation. Addressing the differences in frequency band sensitivity between CNNs and the human visual system (mid-frequency sensitivity with low-frequency sensitivity surpassing high-frequency), we experimentally quantified the CNN contrast sensitivity function and proposed a wavelet adaptive spectrum fusion (WASF) method inspired by biological vision mechanisms to balance cross-frequency image features. Furthermore, we designed a perception frequency block (PFB) that integrates WASF to enhance frequency-domain feature extraction. Based on this, we developed the FE-UNet model, which employs a SAM2 backbone network and incorporates fine-tuned Hiera-Large modules to ensure segmentation accuracy while improving generalization capability. Experiments demonstrate that FE-UNet achieves state-of-the-art performance in cross-domain tasks such as marine organism segmentation and polyp segmentation, showcasing robust adaptability and significant application potential. The code will be released soon.

Zhenglun Kong, Peiyan Dong, Xiaolong Ma et al.

Recently, Vision Transformer (ViT) has continuously established new milestones in the computer vision field, while the high computation and memory cost makes its propagation in industrial production difficult. Pruning, a traditional model compression paradigm for hardware efficiency, has been widely applied in various DNN structures. Nevertheless, it stays ambiguous on how to perform exclusive pruning on the ViT structure. Considering three key points: the structural characteristics, the internal data pattern of ViTs, and the related edge device deployment, we leverage the input token sparsity and propose a computation-aware soft pruning framework, which can be set up on vanilla Transformers of both flatten and CNN-type structures, such as Pooling-based ViT (PiT). More concretely, we design a dynamic attention-based multi-head token selector, which is a lightweight module for adaptive instance-wise token selection. We further introduce a soft pruning technique, which integrates the less informative tokens generated by the selector module into a package token that will participate in subsequent calculations rather than being completely discarded. Our framework is bound to the trade-off between accuracy and computation constraints of specific edge devices through our proposed computation-aware training strategy. Experimental results show that our framework significantly reduces the computation cost of ViTs while maintaining comparable performance on image classification. Moreover, our framework can guarantee the identified model to meet resource specifications of mobile devices and FPGA, and even achieve the real-time execution of DeiT-T on mobile platforms. For example, our method reduces the latency of DeiT-T to 26 ms (26%$\sim $41% superior to existing works) on the mobile device with 0.25%$\sim $4% higher top-1 accuracy on ImageNet.

Hao Tang, Xingwei Liu, Kun Han et al.

Multi-organ segmentation is one of most successful applications of deep learning in medical image analysis. Deep convolutional neural nets (CNNs) have shown great promise in achieving clinically applicable image segmentation performance on CT or MRI images. State-of-the-art CNN segmentation models apply either 2D or 3D convolutions on input images, with pros and cons associated with each method: 2D convolution is fast, less memory-intensive but inadequate for extracting 3D contextual information from volumetric images, while the opposite is true for 3D convolution. To fit a 3D CNN model on CT or MRI images on commodity GPUs, one usually has to either downsample input images or use cropped local regions as inputs, which limits the utility of 3D models for multi-organ segmentation. In this work, we propose a new framework for combining 3D and 2D models, in which the segmentation is realized through high-resolution 2D convolutions, but guided by spatial contextual information extracted from a low-resolution 3D model. We implement a self-attention mechanism to control which 3D features should be used to guide 2D segmentation. Our model is light on memory usage but fully equipped to take 3D contextual information into account. Experiments on multiple organ segmentation datasets demonstrate that by taking advantage of both 2D and 3D models, our method is consistently outperforms existing 2D and 3D models in organ segmentation accuracy, while being able to directly take raw whole-volume image data as inputs.

Bin Duan, Hao Tang, Wei Wang et al.

The major challenge in audio-visual event localization task lies in how to fuse information from multiple modalities effectively. Recent works have shown that attention mechanism is beneficial to the fusion process. In this paper, we propose a novel joint attention mechanism with multimodal fusion methods for audio-visual event localization. Particularly, we present a concise yet valid architecture that effectively learns representations from multiple modalities in a joint manner. Initially, visual features are combined with auditory features and then turned into joint representations. Next, we make use of the joint representations to attend to visual features and auditory features, respectively. With the help of this joint co-attention, new visual and auditory features are produced, and thus both features can enjoy the mutually improved benefits from each other. It is worth noting that the joint co-attention unit is recursive meaning that it can be performed multiple times for obtaining better joint representations progressively. Extensive experiments on the public AVE dataset have shown that the proposed method achieves significantly better results than the state-of-the-art methods.

Xinya Chen, Yanrui Bin, Changxin Gao et al.

Crowd counting is a concerned and challenging task in computer vision. Existing density map based methods excessively focus on the individuals' localization which harms the crowd counting performance in highly congested scenes. In addition, the dependency between the regions of different density is also ignored. In this paper, we propose Relevant Region Prediction (RRP) for crowd counting, which consists of the Count Map and the Region Relation-Aware Module (RRAM). Each pixel in the count map represents the number of heads falling into the corresponding local area in the input image, which discards the detailed spatial information and forces the network pay more attention to counting rather than localizing individuals. Based on the Graph Convolutional Network (GCN), Region Relation-Aware Module is proposed to capture and exploit the important region dependency. The module builds a fully connected directed graph between the regions of different density where each node (region) is represented by weighted global pooled feature, and GCN is learned to map this region graph to a set of relation-aware regions representations. Experimental results on three datasets show that our method obviously outperforms other existing state-of-the-art methods.

Hao Tang, Xiaojuan Qi, Guolei Sun et al.

We propose a novel ECGAN for the challenging semantic image synthesis task. Although considerable improvement has been achieved, the quality of synthesized images is far from satisfactory due to three largely unresolved challenges. 1) The semantic labels do not provide detailed structural information, making it difficult to synthesize local details and structures. 2) The widely adopted CNN operations such as convolution, down-sampling, and normalization usually cause spatial resolution loss and thus cannot fully preserve the original semantic information, leading to semantically inconsistent results. 3) Existing semantic image synthesis methods focus on modeling local semantic information from a single input semantic layout. However, they ignore global semantic information of multiple input semantic layouts, i.e., semantic cross-relations between pixels across different input layouts. To tackle 1), we propose to use edge as an intermediate representation which is further adopted to guide image generation via a proposed attention guided edge transfer module. Edge information is produced by a convolutional generator and introduces detailed structure information. To tackle 2), we design an effective module to selectively highlight class-dependent feature maps according to the original semantic layout to preserve the semantic information. To tackle 3), inspired by current methods in contrastive learning, we propose a novel contrastive learning method, which aims to enforce pixel embeddings belonging to the same semantic class to generate more similar image content than those from different classes. Doing so can capture more semantic relations by explicitly exploring the structures of labeled pixels from multiple input semantic layouts. Experiments on three challenging datasets show that our ECGAN achieves significantly better results than state-of-the-art methods.

Sharmistha Jat, Hao Tang, Partha Talukdar et al.

What is the relationship between sentence representations learned by deep recurrent models against those encoded by the brain? Is there any correspondence between hidden layers of these recurrent models and brain regions when processing sentences? Can these deep models be used to synthesize brain data which can then be utilized in other extrinsic tasks? We investigate these questions using sentences with simple syntax and semantics (e.g., The bone was eaten by the dog.). We consider multiple neural network architectures, including recently proposed ELMo and BERT. We use magnetoencephalography (MEG) brain recording data collected from human subjects when they were reading these simple sentences. Overall, we find that BERT's activations correlate the best with MEG brain data. We also find that the deep network representation can be used to generate brain data from new sentences to augment existing brain data. To the best of our knowledge, this is the first work showing that the MEG brain recording when reading a word in a sentence can be used to distinguish earlier words in the sentence. Our exploration is also the first to use deep neural network representations to generate synthetic brain data and to show that it helps in improving subsequent stimuli decoding task accuracy.

Hao Tang, Xingwei Liu, Xiaohui Xie

Pulmonary nodule detection using low-dose Computed Tomography (CT) is often the first step in lung disease screening and diagnosis. Recently, algorithms based on deep convolutional neural nets have shown great promise for automated nodule detection. Most of the existing deep learning nodule detection systems are constructed in two steps: a) nodule candidates screening and b) false positive reduction, using two different models trained separately. Although it is commonly adopted, the two-step approach not only imposes significant resource overhead on training two independent deep learning models, but also is sub-optimal because it prevents cross-talk between the two. In this work, we present an end-to-end framework for nodule detection, integrating nodule candidate screening and false positive reduction into one model, trained jointly. We demonstrate that the end-to-end system improves the performance by 3.88\% over the two-step approach, while at the same time reducing model complexity by one third and cutting inference time by 3.6 fold. Code will be made publicly available.

Hao Tang, Daniel R. Kim, Xiaohui Xie

Early detection of pulmonary nodules in computed tomography (CT) images is essential for successful outcomes among lung cancer patients. Much attention has been given to deep convolutional neural network (DCNN)-based approaches to this task, but models have relied at least partly on 2D or 2.5D components for inherently 3D data. In this paper, we introduce a novel DCNN approach, consisting of two stages, that is fully three-dimensional end-to-end and utilizes the state-of-the-art in object detection. First, nodule candidates are identified with a U-Net-inspired 3D Faster R-CNN trained using online hard negative mining. Second, false positive reduction is performed by 3D DCNN classifiers trained on difficult examples produced during candidate screening. Finally, we introduce a method to ensemble models from both stages via consensus to give the final predictions. By using this framework, we ranked first of 2887 teams in Season One of Alibaba's 2017 TianChi AI Competition for Healthcare.

Greg Olmschenk, Hao Tang, Zhigang Zhu

Gatherings of thousands to millions of people frequently occur for an enormous variety of events, and automated counting of these high-density crowds is useful for safety, management, and measuring significance of an event. In this work, we show that the regularly accepted labeling scheme of crowd density maps for training deep neural networks is less effective than our alternative inverse k-nearest neighbor (i$k$NN) maps, even when used directly in existing state-of-the-art network structures. We also provide a new network architecture MUD-i$k$NN, which uses multi-scale upsampling via transposed convolutions to take full advantage of the provided i$k$NN labeling. This upsampling combined with the i$k$NN maps further improves crowd counting accuracy. Our new network architecture performs favorably in comparison with the state-of-the-art. However, our labeling and upsampling techniques are generally applicable to existing crowd counting architectures.

Greg Olmschenk, Zhigang Zhu, Hao Tang

In this work, we generalize semi-supervised generative adversarial networks (GANs) from classification problems to regression problems. In the last few years, the importance of improving the training of neural networks using semi-supervised training has been demonstrated for classification problems. We present a novel loss function, called feature contrasting, resulting in a discriminator which can distinguish between fake and real data based on feature statistics. This method avoids potential biases and limitations of alternative approaches. The generalization of semi-supervised GANs to the regime of regression problems of opens their use to countless applications as well as providing an avenue for a deeper understanding of how GANs function. We first demonstrate the capabilities of semi-supervised regression GANs on a toy dataset which allows for a detailed understanding of how they operate in various circumstances. This toy dataset is used to provide a theoretical basis of the semi-supervised regression GAN. We then apply the semi-supervised regression GANs to a number of real-world computer vision applications: age estimation, driving steering angle prediction, and crowd counting from single images. We perform extensive tests of what accuracy can be achieved with significantly reduced annotated data. Through the combination of the theoretical example and real-world scenarios, we demonstrate how semi-supervised GANs can be generalized to regression problems.