Haoji Hu

Jianhong Bai, Zuozhu Liu, Hualiang Wang et al.

Though Self-supervised learning (SSL) has been widely studied as a promising technique for representation learning, it doesn't generalize well on long-tailed datasets due to the majority classes dominating the feature space. Recent work shows that the long-tailed learning performance could be boosted by sampling extra in-domain (ID) data for self-supervised training, however, large-scale ID data which can rebalance the minority classes are expensive to collect. In this paper, we propose an alternative but easy-to-use and effective solution, Contrastive with Out-of-distribution (OOD) data for Long-Tail learning (COLT), which can effectively exploit OOD data to dynamically re-balance the feature space. We empirically identify the counter-intuitive usefulness of OOD samples in SSL long-tailed learning and principally design a novel SSL method. Concretely, we first localize the `head' and `tail' samples by assigning a tailness score to each OOD sample based on its neighborhoods in the feature space. Then, we propose an online OOD sampling strategy to dynamically re-balance the feature space. Finally, we enforce the model to be capable of distinguishing ID and OOD samples by a distribution-level supervised contrastive loss. Extensive experiments are conducted on various datasets and several state-of-the-art SSL frameworks to verify the effectiveness of the proposed method. The results show that our method significantly improves the performance of SSL on long-tailed datasets by a large margin, and even outperforms previous work which uses external ID data. Our code is available at https://github.com/JianhongBai/COLT.

Hualiang Wang, Siming Fu, Xiaoxuan He et al.

Long-tailed learning aims to tackle the crucial challenge that head classes dominate the training procedure under severe class imbalance in real-world scenarios. However, little attention has been given to how to quantify the dominance severity of head classes in the representation space. Motivated by this, we generalize the cosine-based classifiers to a von Mises-Fisher (vMF) mixture model, denoted as vMF classifier, which enables to quantitatively measure representation quality upon the hyper-sphere space via calculating distribution overlap coefficient. To our knowledge, this is the first work to measure representation quality of classifiers and features from the perspective of distribution overlap coefficient. On top of it, we formulate the inter-class discrepancy and class-feature consistency loss terms to alleviate the interference among the classifier weights and align features with classifier weights. Furthermore, a novel post-training calibration algorithm is devised to zero-costly boost the performance via inter-class overlap coefficients. Our method outperforms previous work with a large margin and achieves state-of-the-art performance on long-tailed image classification, semantic segmentation, and instance segmentation tasks (e.g., we achieve 55.0\% overall accuracy with ResNetXt-50 in ImageNet-LT). Our code is available at https://github.com/VipaiLab/vMF\_OP.

Jin Hao, Jiaxiang Liu, Jin Li et al.

A critical step in virtual dental treatment planning is to accurately delineate all tooth-bone structures from CBCT with high fidelity and accurate anatomical information. Previous studies have established several methods for CBCT segmentation using deep learning. However, the inherent resolution discrepancy of CBCT and the loss of occlusal and dentition information largely limited its clinical applicability. Here, we present a Deep Dental Multimodal Analysis (DDMA) framework consisting of a CBCT segmentation model, an intraoral scan (IOS) segmentation model (the most accurate digital dental model), and a fusion model to generate 3D fused crown-root-bone structures with high fidelity and accurate occlusal and dentition information. Our model was trained with a large-scale dataset with 503 CBCT and 28,559 IOS meshes manually annotated by experienced human experts. For CBCT segmentation, we use a five-fold cross validation test, each with 50 CBCT, and our model achieves an average Dice coefficient and IoU of 93.99% and 88.68%, respectively, significantly outperforming the baselines. For IOS segmentations, our model achieves an mIoU of 93.07% and 95.70% on the maxillary and mandible on a test set of 200 IOS meshes, which are 1.77% and 3.52% higher than the state-of-art method. Our DDMA framework takes about 20 to 25 minutes to generate the fused 3D mesh model following the sequential processing order, compared to over 5 hours by human experts. Notably, our framework has been incorporated into a software by a clear aligner manufacturer, and real-world clinical cases demonstrate that our model can visualize crown-root-bone structures during the entire orthodontic treatment and can predict risks like dehiscence and fenestration. These findings demonstrate the potential of multi-modal deep learning to improve the quality of digital dental models and help dentists make better clinical decisions.

Jianhong Bai, Zuozhu Liu, Hualiang Wang et al.

Data imbalance and open-ended distribution are two intrinsic characteristics of the real visual world. Though encouraging progress has been made in tackling each challenge separately, few works dedicated to combining them towards real-world scenarios. While several previous works have focused on classifying close-set samples and detecting open-set samples during testing, it's still essential to be able to classify unknown subjects as human beings. In this paper, we formally define a more realistic task as distribution-agnostic generalized category discovery (DA-GCD): generating fine-grained predictions for both close- and open-set classes in a long-tailed open-world setting. To tackle the challenging problem, we propose a Self-Balanced Co-Advice contrastive framework (BaCon), which consists of a contrastive-learning branch and a pseudo-labeling branch, working collaboratively to provide interactive supervision to resolve the DA-GCD task. In particular, the contrastive-learning branch provides reliable distribution estimation to regularize the predictions of the pseudo-labeling branch, which in turn guides contrastive learning through self-balanced knowledge transfer and a proposed novel contrastive loss. We compare BaCon with state-of-the-art methods from two closely related fields: imbalanced semi-supervised learning and generalized category discovery. The effectiveness of BaCon is demonstrated with superior performance over all baselines and comprehensive analysis across various datasets. Our code is publicly available.

Xiaoxuan He, Yifan Yang, Xinyang Jiang et al.

Vision-Language Pre-training (VLP) has shown the merits of analysing medical images, by leveraging the semantic congruence between medical images and their corresponding reports. It efficiently learns visual representations, which in turn facilitates enhanced analysis and interpretation of intricate imaging data. However, such observation is predominantly justified on single-modality data (mostly 2D images like X-rays), adapting VLP to learning unified representations for medical images in real scenario remains an open challenge. This arises from medical images often encompass a variety of modalities, especially modalities with different various number of dimensions (e.g., 3D images like Computed Tomography). To overcome the aforementioned challenges, we propose an Unified Medical Image Pre-training framework, namely UniMedI, which utilizes diagnostic reports as common semantic space to create unified representations for diverse modalities of medical images (especially for 2D and 3D images). Under the text's guidance, we effectively uncover visual modality information, identifying the affected areas in 2D X-rays and slices containing lesion in sophisticated 3D CT scans, ultimately enhancing the consistency across various medical imaging modalities. To demonstrate the effectiveness and versatility of UniMedI, we evaluate its performance on both 2D and 3D images across 10 different datasets, covering a wide range of medical image tasks such as classification, segmentation, and retrieval. UniMedI has demonstrated superior performance in downstream tasks, showcasing its effectiveness in establishing a universal medical visual representation.

Jianhong Bai, Yuchen Yang, Huanpeng Chu et al.

Quantization has emerged as a promising direction for model compression. Recently, data-free quantization has been widely studied as a promising method to avoid privacy concerns, which synthesizes images as an alternative to real training data. Existing methods use classification loss to ensure the reliability of the synthesized images. Unfortunately, even if these images are well-classified by the pre-trained model, they still suffer from low semantics and homogenization issues. Intuitively, these low-semantic images are sensitive to perturbations, and the pre-trained model tends to have inconsistent output when the generator synthesizes an image with poor semantics. To this end, we propose Robustness-Guided Image Synthesis (RIS), a simple but effective method to enrich the semantics of synthetic images and improve image diversity, further boosting the performance of downstream data-free compression tasks. Concretely, we first introduce perturbations on input and model weight, then define the inconsistency metrics at feature and prediction levels before and after perturbations. On the basis of inconsistency on two levels, we design a robustness optimization objective to enhance the semantics of synthetic images. Moreover, we also make our approach diversity-aware by forcing the generator to synthesize images with small correlations in the label space. With RIS, we achieve state-of-the-art performance for various settings on data-free quantization and can be extended to other data-free compression tasks.

Jianhong Bai, Tianyu He, Yuchi Wang et al. · pku

Recent advances in text-guided video editing have showcased promising results in appearance editing (e.g., stylization). However, video motion editing in the temporal dimension (e.g., from eating to waving), which distinguishes video editing from image editing, is underexplored. In this work, we present UniEdit, a tuning-free framework that supports both video motion and appearance editing by harnessing the power of a pre-trained text-to-video generator within an inversion-then-generation framework. To realize motion editing while preserving source video content, based on the insights that temporal and spatial self-attention layers encode inter-frame and intra-frame dependency respectively, we introduce auxiliary motion-reference and reconstruction branches to produce text-guided motion and source features respectively. The obtained features are then injected into the main editing path via temporal and spatial self-attention layers. Extensive experiments demonstrate that UniEdit covers video motion editing and various appearance editing scenarios, and surpasses the state-of-the-art methods. Our code will be publicly available.

Jianhong Bai, Xiaoshi Wu, Xintao Wang et al.

State-of-the-art video generative models typically learn the distribution of video latents in the VAE space and map them to pixels using a VAE decoder. While this approach can generate high-quality videos, it suffers from slow convergence and is computationally expensive when generating long videos. In this paper, we introduce SemanticGen, a novel solution to address these limitations by generating videos in the semantic space. Our main insight is that, due to the inherent redundancy in videos, the generation process should begin in a compact, high-level semantic space for global planning, followed by the addition of high-frequency details, rather than directly modeling a vast set of low-level video tokens using bi-directional attention. SemanticGen adopts a two-stage generation process. In the first stage, a diffusion model generates compact semantic video features, which define the global layout of the video. In the second stage, another diffusion model generates VAE latents conditioned on these semantic features to produce the final output. We observe that generation in the semantic space leads to faster convergence compared to the VAE latent space. Our method is also effective and computationally efficient when extended to long video generation. Extensive experiments demonstrate that SemanticGen produces high-quality videos and outperforms state-of-the-art approaches and strong baselines.

Jiedong Zhuang, Jiaqi Hu, Lianrui Mu et al.

CLIP has achieved impressive zero-shot performance after pre-training on a large-scale dataset consisting of paired image-text data. Previous works have utilized CLIP by incorporating manually designed visual prompts like colored circles and blur masks into the images to guide the model's attention, showing enhanced zero-shot performance in downstream tasks. Although these methods have achieved promising results, they inevitably alter the original information of the images, which can lead to failure in specific tasks. We propose a train-free method Foveal-Attention CLIP (FALIP), which adjusts the CLIP's attention by inserting foveal attention masks into the multi-head self-attention module. We demonstrate FALIP effectively boosts CLIP zero-shot performance in tasks such as referring expressions comprehension, image classification, and 3D point cloud recognition. Experimental results further show that FALIP outperforms existing methods on most metrics and can augment current methods to enhance their performance.

Xiaoyu Liang, Jiayuan Yu, Lianrui Mu et al.

Although Visual-Language Models (VLMs) have shown impressive capabilities in tasks like visual question answering and image captioning, they still struggle with hallucinations. Analysis of attention distribution in these models shows that VLMs tend to processing textual tokens rather than visual tokens. This imbalance of attention distribution causes VLMs to favor textual knowledge in the case of multimodal knowledge conflicts, resulting in differences from the image information. In this paper, we propose Re-Balancing Contrastive Decoding (RBD) method, which employs textual and visual branches to recalibrate attention distribution in VLMs. Specifically, the textual branch injects image noise to stimulate the model's dependency on text, thereby reducing textual bias. Concurrently, the visual branch focuses on the selection of significant tokens, refining the attention mechanism to highlight the primary subject. This dual-branch strategy enables the RBD method to diminish textual bias while enhancing visual information. Experimental results demonstrate that our method, RBD, outperforms the existing methods by the CHAIR and POPE metrics, mitigate hallucinations without reducing the model's general capabilities.

Lianrui Mu, Jianhong Bai, Xiaoxuan He et al.

Enhancing the domain generalization performance of Face Anti-Spoofing (FAS) techniques has emerged as a research focus. Existing methods are dedicated to extracting domain-invariant features from various training domains. Despite the promising performance, the extracted features inevitably contain residual style feature bias (e.g., illumination, capture device), resulting in inferior generalization performance. In this paper, we propose an alternative and effective solution, the Textually Guided Domain Generalization (TeG-DG) framework, which can effectively leverage text information for cross-domain alignment. Our core insight is that text, as a more abstract and universal form of expression, can capture the commonalities and essential characteristics across various attacks, bridging the gap between different image domains. Contrary to existing vision-language models, the proposed framework is elaborately designed to enhance the domain generalization ability of the FAS task. Concretely, we first design a Hierarchical Attention Fusion (HAF) module to enable adaptive aggregation of visual features at different levels; Then, a Textual-Enhanced Visual Discriminator (TEVD) is proposed for not only better alignment between the two modalities but also to regularize the classifier with unbiased text features. TeG-DG significantly outperforms previous approaches, especially in situations with extremely limited source domain data (~14% and ~12% improvements on HTER and AUC respectively), showcasing impressive few-shot performance.

Huan Wang, Yijun Li, Yuehai Wang et al.

Universal style transfer methods typically leverage rich representations from deep Convolutional Neural Network (CNN) models (e.g., VGG-19) pre-trained on large collections of images. Despite the effectiveness, its application is heavily constrained by the large model size to handle ultra-resolution images given limited memory. In this work, we present a new knowledge distillation method (named Collaborative Distillation) for encoder-decoder based neural style transfer to reduce the convolutional filters. The main idea is underpinned by a finding that the encoder-decoder pairs construct an exclusive collaborative relationship, which is regarded as a new kind of knowledge for style transfer models. Moreover, to overcome the feature size mismatch when applying collaborative distillation, a linear embedding loss is introduced to drive the student network to learn a linear embedding of the teacher's features. Extensive experiments show the effectiveness of our method when applied to different universal style transfer approaches (WCT and AdaIN), even if the model size is reduced by 15.5 times. Especially, on WCT with the compressed models, we achieve ultra-resolution (over 40 megapixels) universal style transfer on a 12GB GPU for the first time. Further experiments on optimization-based stylization scheme show the generality of our algorithm on different stylization paradigms. Our code and trained models are available at https://github.com/mingsun-tse/collaborative-distillation.

Huan Wang, Qiming Zhang, Yuehai Wang et al.

Parameter pruning is a promising approach for CNN compression and acceleration by eliminating redundant model parameters with tolerable performance degrade. Despite its effectiveness, existing regularization-based parameter pruning methods usually drive weights towards zero with large and constant regularization factors, which neglects the fragility of the expressiveness of CNNs, and thus calls for a more gentle regularization scheme so that the networks can adapt during pruning. To achieve this, we propose a new and novel regularization-based pruning method, named IncReg, to incrementally assign different regularization factors to different weights based on their relative importance. Empirical analysis on CIFAR-10 dataset verifies the merits of IncReg. Further extensive experiments with popular CNNs on CIFAR-10 and ImageNet datasets show that IncReg achieves comparable to even better results compared with state-of-the-arts. Our source codes and trained models are available here: https://github.com/mingsun-tse/caffe_increg.

Jiedong Zhuang, Lu Lu, Ming Dai et al.

Multimodal large language models (MLLMs) enhance their perceptual capabilities by integrating visual and textual information. However, processing the massive number of visual tokens incurs a significant computational cost. Existing analysis of the MLLM attention mechanisms remains shallow, leading to coarse-grain token pruning strategies that fail to effectively balance speed and accuracy. In this paper, we conduct a comprehensive investigation of MLLM attention mechanisms with LLaVA. We find that numerous visual tokens and partial attention computations are redundant during the decoding process. Based on this insight, we propose Spatial-Temporal Visual Token Trimming ($\textbf{ST}^{3}$), a framework designed to accelerate MLLM inference without retraining. $\textbf{ST}^{3}$ consists of two primary components: 1) Progressive Visual Token Pruning (\textbf{PVTP}), which eliminates inattentive visual tokens across layers, and 2) Visual Token Annealing (\textbf{VTA}), which dynamically reduces the number of visual tokens in each layer as the generated tokens grow. Together, these techniques deliver around $\mathbf{2\times}$ faster inference with only about $\mathbf{30\%}$ KV cache memory compared to the original LLaVA, while maintaining consistent performance across various datasets. Crucially, $\textbf{ST}^{3}$ can be seamlessly integrated into existing pre-trained MLLMs, providing a plug-and-play solution for efficient inference.

Xiaoyu Liang, Chaofeng Guan, Jiaying Lu et al.

Vision-Language Models (VLMs) have achieved notable success in multimodal tasks but face practical limitations due to the quadratic complexity of decoder attention mechanisms and autoregressive generation. Existing methods like FASTV and VTW have achieved notable results in reducing redundant visual tokens, but these approaches focus on pruning tokens in a single forward pass without systematically analyzing the redundancy of visual tokens throughout the entire generation process. In this paper, we introduce a dynamic pruning strategy tailored for VLMs, namedDynamic Rate (DyRate), which progressively adjusts the compression rate during generation. Our analysis of the distribution of attention reveals that the importance of visual tokens decreases throughout the generation process, inspiring us to adopt a more aggressive compression rate. By integrating a lightweight predictor based on attention distribution, our approach enables flexible adjustment of pruning rates based on the attention distribution. Our experimental results demonstrate that our method not only reduces computational demands but also maintains the quality of responses.

Lianrui Mu, Xingze Zhou, Wenjie Zheng et al.

Creating realistic pose-guided image-to-video character animations while preserving facial identity remains challenging, especially in complex and dynamic scenarios such as dancing, where precise identity consistency is crucial. Existing methods frequently encounter difficulties maintaining facial coherence due to misalignments between facial landmarks extracted from driving videos that provide head pose and expression cues and the facial geometry of the reference images. To address this limitation, we introduce the Facial Landmarks Transformation (FLT) method, which leverages a 3D Morphable Model to address this limitation. FLT converts 2D landmarks into a 3D face model, adjusts the 3D face model to align with the reference identity, and then transforms them back into 2D landmarks to guide the image-to-video generation process. This approach ensures accurate alignment with the reference facial geometry, enhancing the consistency between generated videos and reference images. Experimental results demonstrate that FLT effectively preserves facial identity, significantly improving pose-guided character animation models.

Chen Lin, Zheyang Li, Bo Peng et al.

This paper introduces a post-training quantization~(PTQ) method achieving highly efficient Convolutional Neural Network~ (CNN) quantization with high performance. Previous PTQ methods usually reduce compression error via performing layer-by-layer parameters calibration. However, with lower representational ability of extremely compressed parameters (e.g., the bit-width goes less than 4), it is hard to eliminate all the layer-wise errors. This work addresses this issue via proposing a unit-wise feature reconstruction algorithm based on an observation of second order Taylor series expansion of the unit-wise error. It indicates that leveraging the interaction between adjacent layers' parameters could compensate layer-wise errors better. In this paper, we define several adjacent layers as a Basic-Unit, and present a unit-wise post-training algorithm which can minimize quantization error. This method achieves near-original accuracy on ImageNet and COCO when quantizing FP32 models to INT4 and INT3.

Kuan Zhang, Haoji Hu, Kenneth Philbrick et al.

There is a growing demand for high-resolution (HR) medical images in both the clinical and research applications. Image quality is inevitably traded off with the acquisition time for better patient comfort, lower examination costs, dose, and fewer motion-induced artifacts. For many image-based tasks, increasing the apparent resolution in the perpendicular plane to produce multi-planar reformats or 3D images is commonly used. Single image super-resolution (SR) is a promising technique to provide HR images based on unsupervised learning to increase resolution of a 2D image, but there are few reports on 3D SR. Further, perceptual loss is proposed in the literature to better capture the textual details and edges than using pixel-wise loss functions, by comparing the semantic distances in the high-dimensional feature space of a pre-trained 2D network (e.g., VGG). However, it is not clear how one should generalize it to 3D medical images, and the attendant implications are still unclear. In this paper, we propose a framework called SOUP-GAN: Super-resolution Optimized Using Perceptual-tuned Generative Adversarial Network (GAN), in order to produce thinner slice (e.g., high resolution in the 'Z' plane) medical images with anti-aliasing and deblurring. The proposed method outperforms other conventional resolution-enhancement methods and previous SR work on medical images upon both qualitative and quantitative comparisons. Specifically, we examine the model in terms of its generalization for various SR ratios and imaging modalities. By addressing those limitations, our model shows promise as a novel 3D SR interpolation technique, providing potential applications in both clinical and research settings.

Bianjiang Yang, Zi Hui, Haoji Hu et al.

Although the facial makeup transfer network has achieved high-quality performance in generating perceptually pleasing makeup images, its capability is still restricted by the massive computation and storage of the network architecture. We address this issue by compressing facial makeup transfer networks with collaborative distillation and kernel decomposition. The main idea of collaborative distillation is underpinned by a finding that the encoder-decoder pairs construct an exclusive collaborative relationship, which is regarded as a new kind of knowledge for low-level vision tasks. For kernel decomposition, we apply the depth-wise separation of convolutional kernels to build a light-weighted Convolutional Neural Network (CNN) from the original network. Extensive experiments show the effectiveness of the compression method when applied to the state-of-the-art facial makeup transfer network -- BeautyGAN.

Huan Wang, Qiming Zhang, Yuehai Wang et al.

Parameter pruning is a promising approach for CNN compression and acceleration by eliminating redundant model parameters with tolerable performance loss. Despite its effectiveness, existing regularization-based parameter pruning methods usually drive weights towards zero with large and constant regularization factors, which neglects the fact that the expressiveness of CNNs is fragile and needs a more gentle way of regularization for the networks to adapt during pruning. To solve this problem, we propose a new regularization-based pruning method (named IncReg) to incrementally assign different regularization factors to different weight groups based on their relative importance, whose effectiveness is proved on popular CNNs compared with state-of-the-art methods.

Yuxin Zhang, Huan Wang, Yang Luo et al.

Despite enjoying extensive applications in video analysis, three-dimensional convolutional neural networks (3D CNNs)are restricted by their massive computation and storage consumption. To solve this problem, we propose a threedimensional regularization-based neural network pruning method to assign different regularization parameters to different weight groups based on their importance to the network. Further we analyze the redundancy and computation cost for each layer to determine the different pruning ratios. Experiments show that pruning based on our method can lead to 2x theoretical speedup with only 0.41% accuracy loss for 3DResNet18 and 3.28% accuracy loss for C3D. The proposed method performs favorably against other popular methods for model compression and acceleration.