Wei Huang

Wei Huang, Zhiliang Peng, Li Dong et al. · microsoft-research

Large vision Transformers (ViTs) driven by self-supervised pre-training mechanisms achieved unprecedented progress. Lightweight ViT models limited by the model capacity, however, benefit little from those pre-training mechanisms. Knowledge distillation defines a paradigm to transfer representations from large (teacher) models to small (student) ones. However, the conventional single-stage distillation easily gets stuck on task-specific transfer, failing to retain the task-agnostic knowledge crucial for model generalization. In this study, we propose generic-to-specific distillation (G2SD), to tap the potential of small ViT models under the supervision of large models pre-trained by masked autoencoders. In generic distillation, decoder of the small model is encouraged to align feature predictions with hidden representations of the large model, so that task-agnostic knowledge can be transferred. In specific distillation, predictions of the small model are constrained to be consistent with those of the large model, to transfer task-specific features which guarantee task performance. With G2SD, the vanilla ViT-Small model respectively achieves 98.7%, 98.1% and 99.3% the performance of its teacher (ViT-Base) for image classification, object detection, and semantic segmentation, setting a solid baseline for two-stage vision distillation. Code will be available at https://github.com/pengzhiliang/G2SD.

Xiaosong Zhang, Yunjie Tian, Wei Huang et al.

Recently, masked image modeling (MIM) has offered a new methodology of self-supervised pre-training of vision transformers. A key idea of efficient implementation is to discard the masked image patches (or tokens) throughout the target network (encoder), which requires the encoder to be a plain vision transformer (e.g., ViT), albeit hierarchical vision transformers (e.g., Swin Transformer) have potentially better properties in formulating vision inputs. In this paper, we offer a new design of hierarchical vision transformers named HiViT (short for Hierarchical ViT) that enjoys both high efficiency and good performance in MIM. The key is to remove the unnecessary "local inter-unit operations", deriving structurally simple hierarchical vision transformers in which mask-units can be serialized like plain vision transformers. For this purpose, we start with Swin Transformer and (i) set the masking unit size to be the token size in the main stage of Swin Transformer, (ii) switch off inter-unit self-attentions before the main stage, and (iii) eliminate all operations after the main stage. Empirical studies demonstrate the advantageous performance of HiViT in terms of fully-supervised, self-supervised, and transfer learning. In particular, in running MAE on ImageNet-1K, HiViT-B reports a +0.6% accuracy gain over ViT-B and a 1.9$\times$ speed-up over Swin-B, and the performance gain generalizes to downstream tasks of detection and segmentation. Code will be made publicly available.

Yinda Chen, Wei Huang, Xiaoyu Liu et al.

Instance segmentation in electron microscopy (EM) volumes is tough due to complex shapes and sparse annotations. Self-supervised learning helps but still struggles with intricate visual patterns in EM. To address this, we propose a pretraining framework that enhances multiscale consistency in EM volumes. Our approach leverages a Siamese network architecture, integrating both strong and weak data augmentations to effectively extract multiscale features. We uphold voxel-level coherence by reconstructing the original input data from these augmented instances. Furthermore, we incorporate cross-attention mechanisms to facilitate fine-grained feature alignment between these augmentations. Finally, we apply contrastive learning techniques across a feature pyramid, allowing us to distill distinctive representations spanning various scales. After pretraining on four large-scale EM datasets, our framework significantly improves downstream tasks like neuron and mitochondria segmentation, especially with limited finetuning data. It effectively captures voxel and feature consistency, showing promise for learning transferable representations for EM analysis.

Ganglai Wang, Peng Zhang, Lei Xie et al.

Talking face generation with great practical significance has attracted more attention in recent audio-visual studies. How to achieve accurate lip synchronization is a long-standing challenge to be further investigated. Motivated by xxx, in this paper, an AttnWav2Lip model is proposed by incorporating spatial attention module and channel attention module into lip-syncing strategy. Rather than focusing on the unimportant regions of the face image, the proposed AttnWav2Lip model is able to pay more attention on the lip region reconstruction. To our limited knowledge, this is the first attempt to introduce attention mechanism to the scheme of talking face generation. An extensive experiments have been conducted to evaluate the effectiveness of the proposed model. Compared to the baseline measured by LSE-D and LSE-C metrics, a superior performance has been demonstrated on the benchmark lip synthesis datasets, including LRW, LRS2 and LRS3.

Cunhang Fan, Hongyu Zhang, Wei Huang et al.

Auditory Attention Detection (AAD) aims to detect target speaker from brain signals in a multi-speaker environment. Although EEG-based AAD methods have shown promising results in recent years, current approaches primarily rely on traditional convolutional neural network designed for processing Euclidean data like images. This makes it challenging to handle EEG signals, which possess non-Euclidean characteristics. In order to address this problem, this paper proposes a dynamical graph self-distillation (DGSD) approach for AAD, which does not require speech stimuli as input. Specifically, to effectively represent the non-Euclidean properties of EEG signals, dynamical graph convolutional networks are applied to represent the graph structure of EEG signals, which can also extract crucial features related to auditory spatial attention in EEG signals. In addition, to further improve AAD detection performance, self-distillation, consisting of feature distillation and hierarchical distillation strategies at each layer, is integrated. These strategies leverage features and classification results from the deepest network layers to guide the learning of shallow layers. Our experiments are conducted on two publicly available datasets, KUL and DTU. Under a 1-second time window, we achieve results of 90.0\% and 79.6\% accuracy on KUL and DTU, respectively. We compare our DGSD method with competitive baselines, and the experimental results indicate that the detection performance of our proposed DGSD method is not only superior to the best reproducible baseline but also significantly reduces the number of trainable parameters by approximately 100 times.

Songhao Han, Wei Huang, Hairong Shi et al.

The advancement of Large Vision Language Models (LVLMs) has significantly improved multimodal understanding, yet challenges remain in video reasoning tasks due to the scarcity of high-quality, large-scale datasets. Existing video question-answering (VideoQA) datasets often rely on costly manual annotations with insufficient granularity or automatic construction methods with redundant frame-by-frame analysis, limiting their scalability and effectiveness for complex reasoning. To address these challenges, we introduce VideoEspresso, a novel dataset that features VideoQA pairs preserving essential spatial details and temporal coherence, along with multimodal annotations of intermediate reasoning steps. Our construction pipeline employs a semantic-aware method to reduce redundancy, followed by generating QA pairs using GPT-4o. We further develop video Chain-of-Thought (CoT) annotations to enrich reasoning processes, guiding GPT-4o in extracting logical relationships from QA pairs and video content. To exploit the potential of high-quality VideoQA pairs, we propose a Hybrid LVLMs Collaboration framework, featuring a Frame Selector and a two-stage instruction fine-tuned reasoning LVLM. This framework adaptively selects core frames and performs CoT reasoning using multimodal evidence. Evaluated on our proposed benchmark with 14 tasks against 9 popular LVLMs, our method outperforms existing baselines on most tasks, demonstrating superior video reasoning capabilities. Our code and dataset will be released at: https://github.com/hshjerry/VideoEspresso

Wuyang Chen, Wei Huang, Xianzhi Du et al.

This work targets automated designing and scaling of Vision Transformers (ViTs). The motivation comes from two pain spots: 1) the lack of efficient and principled methods for designing and scaling ViTs; 2) the tremendous computational cost of training ViT that is much heavier than its convolution counterpart. To tackle these issues, we propose As-ViT, an auto-scaling framework for ViTs without training, which automatically discovers and scales up ViTs in an efficient and principled manner. Specifically, we first design a "seed" ViT topology by leveraging a training-free search process. This extremely fast search is fulfilled by a comprehensive study of ViT's network complexity, yielding a strong Kendall-tau correlation with ground-truth accuracies. Second, starting from the "seed" topology, we automate the scaling rule for ViTs by growing widths/depths to different ViT layers. This results in a series of architectures with different numbers of parameters in a single run. Finally, based on the observation that ViTs can tolerate coarse tokenization in early training stages, we propose a progressive tokenization strategy to train ViTs faster and cheaper. As a unified framework, As-ViT achieves strong performance on classification (83.5% top1 on ImageNet-1k) and detection (52.7% mAP on COCO) without any manual crafting nor scaling of ViT architectures: the end-to-end model design and scaling process cost only 12 hours on one V100 GPU. Our code is available at https://github.com/VITA-Group/AsViT.

Zhiliang Peng, Wei Huang, Shanzhi Gu et al.

Within Convolutional Neural Network (CNN), the convolution operations are good at extracting local features but experience difficulty to capture global representations. Within visual transformer, the cascaded self-attention modules can capture long-distance feature dependencies but unfortunately deteriorate local feature details. In this paper, we propose a hybrid network structure, termed Conformer, to take advantage of convolutional operations and self-attention mechanisms for enhanced representation learning. Conformer roots in the Feature Coupling Unit (FCU), which fuses local features and global representations under different resolutions in an interactive fashion. Conformer adopts a concurrent structure so that local features and global representations are retained to the maximum extent. Experiments show that Conformer, under the comparable parameter complexity, outperforms the visual transformer (DeiT-B) by 2.3% on ImageNet. On MSCOCO, it outperforms ResNet-101 by 3.7% and 3.6% mAPs for object detection and instance segmentation, respectively, demonstrating the great potential to be a general backbone network. Code is available at https://github.com/pengzhiliang/Conformer.

Xinwei Yuan, Shu Han, Wei Huang et al.

Deep learning based intrusion detection systems (DL-based IDS) have emerged as one of the best choices for providing security solutions against various network intrusion attacks. However, due to the emergence and development of adversarial deep learning technologies, it becomes challenging for the adoption of DL models into IDS. In this paper, we propose a novel IDS architecture that can enhance the robustness of IDS against adversarial attacks by combining conventional machine learning (ML) models and Deep Learning models. The proposed DLL-IDS consists of three components: DL-based IDS, adversarial example (AE) detector, and ML-based IDS. We first develop a novel AE detector based on the local intrinsic dimensionality (LID). Then, we exploit the low attack transferability between DL models and ML models to find a robust ML model that can assist us in determining the maliciousness of AEs. If the input traffic is detected as an AE, the ML-based IDS will predict the maliciousness of input traffic, otherwise the DL-based IDS will work for the prediction. The fusion mechanism can leverage the high prediction accuracy of DL models and low attack transferability between DL models and ML models to improve the robustness of the whole system. In our experiments, we observe a significant improvement in the prediction performance of the IDS when subjected to adversarial attack, achieving high accuracy with low resource consumption.

Yujin Han, Andi Han, Wei Huang et al.

Despite the remarkable success of diffusion models (DMs) in data generation, they exhibit specific failure cases with unsatisfactory outputs. We focus on one such limitation: the ability of DMs to learn hidden rules between image features. Specifically, for image data with dependent features ($\mathbf{x}$) and ($\mathbf{y}$) (e.g., the height of the sun ($\mathbf{x}$) and the length of the shadow ($\mathbf{y}$)), we investigate whether DMs can accurately capture the inter-feature rule ($p(\mathbf{y}|\mathbf{x})$). Empirical evaluations on mainstream DMs (e.g., Stable Diffusion 3.5) reveal consistent failures, such as inconsistent lighting-shadow relationships and mismatched object-mirror reflections. Inspired by these findings, we design four synthetic tasks with strongly correlated features to assess DMs' rule-learning abilities. Extensive experiments show that while DMs can identify coarse-grained rules, they struggle with fine-grained ones. Our theoretical analysis demonstrates that DMs trained via denoising score matching (DSM) exhibit constant errors in learning hidden rules, as the DSM objective is not compatible with rule conformity. To mitigate this, we introduce a common technique - incorporating additional classifier guidance during sampling, which achieves (limited) improvements. Our analysis reveals that the subtle signals of fine-grained rules are challenging for the classifier to capture, providing insights for future exploration.

Xue Xia, Daiwei Zhang, Wenxuan Song et al.

Automated feature detection in historical maps can significantly accelerate the reconstruction of the geospatial past. However, this process is often constrained by the time-consuming task of manually digitizing sufficient high-quality training data. The emergence of visual foundation models, such as the Segment Anything Model (SAM), offers a promising solution due to their remarkable generalization capabilities and rapid adaptation to new data distributions. Despite this, directly applying SAM in a zero-shot manner to historical map segmentation poses significant challenges, including poor recognition of certain geospatial features and a reliance on input prompts, which limits its ability to be fully automated. To address these challenges, we introduce MapSAM, a parameter-efficient fine-tuning strategy that adapts SAM into a prompt-free and versatile solution for various downstream historical map segmentation tasks. Specifically, we employ Weight-Decomposed Low-Rank Adaptation (DoRA) to integrate domain-specific knowledge into the image encoder. Additionally, we develop an automatic prompt generation process, eliminating the need for manual input. We further enhance the positional prompt in SAM, transforming it into a higher-level positional-semantic prompt, and modify the cross-attention mechanism in the mask decoder with masked attention for more effective feature aggregation. The proposed MapSAM framework demonstrates promising performance across two distinct historical map segmentation tasks: one focused on linear features and the other on areal features. Experimental results show that it adapts well to various features, even when fine-tuned with extremely limited data (e.g. 10 shots).

Zhitong Xiong, Wei Huang, Jingtao Hu et al.

Generating 3D city models rapidly is crucial for many applications. Monocular height estimation is one of the most efficient and timely ways to obtain large-scale geometric information. However, existing works focus primarily on training and testing models using unbiased datasets, which does not align well with real-world applications. Therefore, we propose a new benchmark dataset to study the transferability of height estimation models in a cross-dataset setting. To this end, we first design and construct a large-scale benchmark dataset for cross-dataset transfer learning on the height estimation task. This benchmark dataset includes a newly proposed large-scale synthetic dataset, a newly collected real-world dataset, and four existing datasets from different cities. Next, a new experimental protocol, few-shot cross-dataset transfer, is designed. Furthermore, in this paper, we propose a scale-deformable convolution module to enhance the window-based Transformer for handling the scale-variation problem in the height estimation task. Experimental results have demonstrated the effectiveness of the proposed methods in the traditional and cross-dataset transfer settings. The datasets and codes are publicly available at https://mediatum.ub.tum.de/1662763 and https://thebenchmarkh.github.io/.