Zhihai He

Yi Zhang, Ce Zhang, Xueting Hu et al. · eth-zurich

Recently, large-scale pre-trained vision-language models (e.g. CLIP and ALIGN) have demonstrated remarkable effectiveness in acquiring transferable visual representations. To leverage the valuable knowledge encoded within these models for downstream tasks, several fine-tuning approaches, including prompt tuning methods and adapter-based methods, have been developed to adapt vision-language models effectively with supervision. However, these methods rely on the availability of annotated samples, which can be labor-intensive and time-consuming to acquire, thus limiting scalability. To address this issue, in this work, we design an unsupervised fine-tuning approach for vision-language models called Unsupervised Prototype Adapter (UP-Adapter). Specifically, for the unannotated target datasets, we leverage the text-image aligning capability of CLIP to automatically select the most confident samples for each class. Utilizing these selected samples, we generate class prototypes, which serve as the initialization for the learnable prototype model. After fine-tuning, the prototype model prediction is combined with the original CLIP's prediction by a residual connection to perform downstream recognition tasks. Our extensive experimental results on image recognition and domain generalization show that the proposed unsupervised method outperforms 8-shot CoOp, 8-shot Tip-Adapter, and also the state-of-the-art UPL method by large margins.

Yi Zhang, Ce Zhang, Yushun Tang et al. · cmu

Large-scale pre-trained Vision-Language Models (VLMs), such as CLIP, establish the correlation between texts and images, achieving remarkable success on various downstream tasks with fine-tuning. In existing fine-tuning methods, the class-specific text description is matched against the whole image. We recognize that this whole image matching is not effective since images from the same class often contain a set of different semantic objects, and an object further consists of a set of semantic parts or concepts. Individual semantic parts or concepts may appear in image samples from different classes. To address this issue, in this paper, we develop a new method called cross-model concept learning and inference (CCLI). Using the powerful text-image correlation capability of CLIP, our method automatically learns a large set of distinctive visual concepts from images using a set of semantic text concepts. Based on these visual concepts, we construct a discriminative representation of images and learn a concept inference network to perform downstream image classification tasks, such as few-shot learning and domain generalization. Extensive experimental results demonstrate that our CCLI method is able to improve the performance upon the current state-of-the-art methods by large margins, for example, by up to 8.0% improvement on few-shot learning and by up to 1.3% for domain generalization.

Xueting Hu, Ce Zhang, Yi Zhang et al. · cmu

Pre-trained Vision-Language Models (VLMs), such as CLIP, have shown enhanced performance across a range of tasks that involve the integration of visual and linguistic modalities. When CLIP is used for depth estimation tasks, the patches, divided from the input images, can be combined with a series of semantic descriptions of the depth information to obtain similarity results. The coarse estimation of depth is then achieved by weighting and summing the depth values, called depth bins, corresponding to the predefined semantic descriptions. The zero-shot approach circumvents the computational and time-intensive nature of traditional fully-supervised depth estimation methods. However, this method, utilizing fixed depth bins, may not effectively generalize as images from different scenes may exhibit distinct depth distributions. To address this challenge, we propose a few-shot-based method which learns to adapt the VLMs for monocular depth estimation to balance training costs and generalization capabilities. Specifically, it assigns different depth bins for different scenes, which can be selected by the model during inference. Additionally, we incorporate learnable prompts to preprocess the input text to convert the easily human-understood text into easily model-understood vectors and further enhance the performance. With only one image per scene for training, our extensive experiment results on the NYU V2 and KITTI dataset demonstrate that our method outperforms the previous state-of-the-art method by up to 10.6\% in terms of MARE.

Zhehan Kan, Shuoshuo Chen, Ce Zhang et al. · cmu

A central challenge in human pose estimation, as well as in many other machine learning and prediction tasks, is the generalization problem. The learned network does not have the capability to characterize the prediction error, generate feedback information from the test sample, and correct the prediction error on the fly for each individual test sample, which results in degraded performance in generalization. In this work, we introduce a self-correctable and adaptable inference (SCAI) method to address the generalization challenge of network prediction and use human pose estimation as an example to demonstrate its effectiveness and performance. We learn a correction network to correct the prediction result conditioned by a fitness feedback error. This feedback error is generated by a learned fitness feedback network which maps the prediction result to the original input domain and compares it against the original input. Interestingly, we find that this self-referential feedback error is highly correlated with the actual prediction error. This strong correlation suggests that we can use this error as feedback to guide the correction process. It can be also used as a loss function to quickly adapt and optimize the correction network during the inference process. Our extensive experimental results on human pose estimation demonstrate that the proposed SCAI method is able to significantly improve the generalization capability and performance of human pose estimation.

Ce Zhang, Kailiang Wu, Zhihai He · cmu

Given an unknown dynamical system, what is the minimum number of samples needed for effective learning of its governing laws and accurate prediction of its future evolution behavior, and how to select these critical samples? In this work, we propose to explore this problem based on a design approach. Starting from a small initial set of samples, we adaptively discover critical samples to achieve increasingly accurate learning of the system evolution. One central challenge here is that we do not know the network modeling error since the ground-truth system state is unknown, which is however needed for critical sampling. To address this challenge, we introduce a multi-step reciprocal prediction network where forward and backward evolution networks are designed to learn the temporal evolution behavior in the forward and backward time directions, respectively. Very interestingly, we find that the desired network modeling error is highly correlated with the multi-step reciprocal prediction error, which can be directly computed from the current system state. This allows us to perform a dynamic selection of critical samples from regions with high network modeling errors for dynamical systems. Additionally, a joint spatial-temporal evolution network is introduced which incorporates spatial dynamics modeling into the temporal evolution prediction for robust learning of the system evolution operator with few samples. Our extensive experimental results demonstrate that our proposed method is able to dramatically reduce the number of samples needed for effective learning and accurate prediction of evolution behaviors of unknown dynamical systems by up to hundreds of times.

Yi Zhang, Chun-Wun Cheng, Ke Yu et al.

In this paper, we consider the problem of prototype-based vision-language reasoning problem. We observe that existing methods encounter three major challenges: 1) escalating resource demands and prolonging training times, 2) contending with excessive learnable parameters, and 3) fine-tuning based only on a single modality. These challenges will hinder their capability to adapt Vision-Language Models (VLMs) to downstream tasks. Motivated by this critical observation, we propose a novel method called NODE-Adapter, which utilizes Neural Ordinary Differential Equations for better vision-language reasoning. To fully leverage both visual and textual modalities and estimate class prototypes more effectively and accurately, we divide our method into two stages: cross-modal prototype construction and cross-modal prototype optimization using neural ordinary differential equations. Specifically, we exploit VLM to encode hand-crafted prompts into textual features and few-shot support images into visual features. Then, we estimate the textual prototype and visual prototype by averaging the textual features and visual features, respectively, and adaptively combine the textual prototype and visual prototype to construct the cross-modal prototype. To alleviate the prototype bias, we then model the prototype optimization process as an initial value problem with Neural ODEs to estimate the continuous gradient flow. Our extensive experimental results, which cover few-shot classification, domain generalization, and visual reasoning on human-object interaction, demonstrate that the proposed method significantly outperforms existing state-of-the-art approaches.

Shichao Kan, Zhiquan He, Yigang Cen et al.

Recent methods for deep metric learning have been focusing on designing different contrastive loss functions between positive and negative pairs of samples so that the learned feature embedding is able to pull positive samples of the same class closer and push negative samples from different classes away from each other. In this work, we recognize that there is a significant semantic gap between features at the intermediate feature layer and class labels at the final output layer. To bridge this gap, we develop a contrastive Bayesian analysis to characterize and model the posterior probabilities of image labels conditioned by their features similarity in a contrastive learning setting. This contrastive Bayesian analysis leads to a new loss function for deep metric learning. To improve the generalization capability of the proposed method onto new classes, we further extend the contrastive Bayesian loss with a metric variance constraint. Our experimental results and ablation studies demonstrate that the proposed contrastive Bayesian metric learning method significantly improves the performance of deep metric learning in both supervised and pseudo-supervised scenarios, outperforming existing methods by a large margin.

Yi Zhang, Ke Yu, Siqi Wu et al.

In this paper, we propose Conceptual Codebook Learning (CoCoLe), a novel fine-tuning method for vision-language models (VLMs) to address the challenge of improving the generalization capability of VLMs while fine-tuning them on downstream tasks in a few-shot setting. We recognize that visual concepts, such as textures, shapes, and colors are naturally transferable across domains and play a crucial role in generalization tasks. Motivated by this interesting finding, we learn a conceptual codebook consisting of visual concepts as keys and conceptual prompts as values, which serves as a link between the image encoder's outputs and the text encoder's inputs. Specifically, for a given image, we leverage the codebook to identify the most relevant conceptual prompts associated with the class embeddings to perform the classification. Additionally, we incorporate a handcrafted concept cache as a regularization to alleviate the overfitting issues in low-shot scenarios. We observe that this conceptual codebook learning method is able to achieve enhanced alignment between visual and linguistic modalities. Extensive experimental results demonstrate that our CoCoLe method remarkably outperforms the existing state-of-the-art methods across various evaluation settings, including base-to-new generalization, cross-dataset evaluation, and domain generalization tasks. Detailed ablation studies further confirm the efficacy of each component in CoCoLe.

Yijia Wang, Yiqing Shen, Weiming Chen et al.

Existing image editing methods can handle simple editing instructions very well. To deal with complex editing instructions, they often need to jointly fine-tune the large language models (LLMs) and diffusion models (DMs), which involves very high computational complexity and training cost. To address this issue, we propose a new method, called \textbf{C}omplex \textbf{I}mage \textbf{E}diting via \textbf{L}LM \textbf{R}easoning (CIELR), which converts a complex user instruction into a set of simple and explicit editing actions, eliminating the need for jointly fine-tuning the large language models and diffusion models. Specifically, we first construct a structured semantic representation of the input image using foundation models. Then, we introduce an iterative update mechanism that can progressively refine this representation, obtaining a fine-grained visual representation of the image scene. This allows us to perform complex and flexible image editing tasks. Extensive experiments on the SmartEdit Reasoning Scenario Set show that our method surpasses the previous state-of-the-art by 9.955 dB in PSNR, indicating its superior preservation of regions that should remain consistent. Due to the limited number of samples of public datasets of complex image editing with reasoning, we construct a benchmark named CIEBench, containing 86 image samples, together with a metric specifically for reasoning-based image editing. CIELR also outperforms previous methods on this benchmark. The code and dataset are available at \href{https://github.com/Jia-shao/Reasoning-Editing}{https://github.com/Jia-shao/Reasoning-Editing}.

Weiming Chen, Zhihan Zhu, Yijia Wang et al.

Rectified flow (RF) models have recently demonstrated superior generative performance compared to DDIM-based diffusion models. However, in real-world applications, they suffer from two major challenges: (1) low inversion accuracy that hinders the consistency with the source image, and (2) entangled multimodal attention in diffusion transformers, which hinders precise attention control. To address the first challenge, we propose an efficient high-order inversion method for rectified flow models based on the Runge-Kutta solver of differential equations. To tackle the second challenge, we introduce Decoupled Diffusion Transformer Attention (DDTA), a novel mechanism that disentangles text and image attention inside the multimodal diffusion transformers, enabling more precise semantic control. Extensive experiments on image reconstruction and text-guided editing tasks demonstrate that our method achieves state-of-the-art performance in terms of fidelity and editability. Code is available at https://github.com/wmchen/RKSovler_DDTA.

Yi Zhang, Ce Zhang, Ke Yu et al. · cmu

Contrastive Language-Image Pretraining (CLIP) model has exhibited remarkable efficacy in establishing cross-modal connections between texts and images, yielding impressive performance across a broad spectrum of downstream applications through fine-tuning. However, for generalization tasks, the current fine-tuning methods for CLIP, such as CoOp and CoCoOp, demonstrate relatively low performance on some fine-grained datasets. We recognize the underlying reason is that these previous methods only projected global features into the prompt, neglecting the various visual concepts, such as colors, shapes, and sizes, which are naturally transferable across domains and play a crucial role in generalization tasks. To address this issue, in this work, we propose Concept-Guided Prompt Learning (CPL) for vision-language models. Specifically, we leverage the well-learned knowledge of CLIP to create a visual concept cache to enable concept-guided prompting. In order to refine the text features, we further develop a projector that transforms multi-level visual features into text features. We observe that this concept-guided prompt learning approach is able to achieve enhanced consistency between visual and linguistic modalities. Extensive experimental results demonstrate that our CPL method significantly improves generalization capabilities compared to the current state-of-the-art methods.

Yushun Tang, Shuoshuo Chen, Jiyuan Jia et al.

Fully test-time adaptation aims to adapt a network model online based on sequential analysis of input samples during the inference stage. We observe that, when applying a transformer network model into a new domain, the self-attention profiles of image samples in the target domain deviate significantly from those in the source domain, which results in large performance degradation during domain changes. To address this important issue, we propose a new structure for the self-attention modules in the transformer. Specifically, we incorporate three domain-conditioning vectors, called domain conditioners, into the query, key, and value components of the self-attention module. We learn a network to generate these three domain conditioners from the class token at each transformer network layer. We find that, during fully online test-time adaptation, these domain conditioners at each transform network layer are able to gradually remove the impact of domain shift and largely recover the original self-attention profile. Our extensive experimental results demonstrate that the proposed domain-conditioned transformer significantly improves the online fully test-time domain adaptation performance and outperforms existing state-of-the-art methods by large margins.

Yushun Tang, Ziqiong Liu, Jiyuan Jia et al.

Online test-time adaptation aims to dynamically adjust a network model in real-time based on sequential input samples during the inference stage. In this work, we find that, when applying a transformer network model to a new target domain, the Query, Key, and Value features of its self-attention module often change significantly from those in the source domain, leading to substantial performance degradation of the transformer model. To address this important issue, we propose to develop a new approach to progressively recalibrate the self-attention at each layer using a local linear transform parameterized by conditioned scale and shift factors. We consider the online model adaptation from the source domain to the target domain as a progressive domain shift separation process. At each transformer network layer, we learn a Domain Separation Network to extract the domain shift feature, which is used to predict the scale and shift parameters for self-attention recalibration using a Factor Generator Network. These two lightweight networks are adapted online during inference. Experimental results on benchmark datasets demonstrate that the proposed progressive conditioned scale-shift recalibration (PCSR) method is able to significantly improve the online test-time domain adaptation performance by a large margin of up to 3.9\% in classification accuracy on the ImageNet-C dataset.

Ziqiong Liu, Yushun Tang, Junyang Ji et al.

Test-time adaptation (TTA) refers to adjusting the model during the testing phase to cope with changes in sample distribution and enhance the model's adaptability to new environments. In real-world scenarios, models often encounter samples from unseen (out-of-distribution, OOD) categories. Misclassifying these as known (in-distribution, ID) classes not only degrades predictive accuracy but can also impair the adaptation process, leading to further errors on subsequent ID samples. Many existing TTA methods suffer substantial performance drops under such conditions. To address this challenge, we propose a Hierarchical Ladder Network that extracts OOD features from class tokens aggregated across all Transformer layers. OOD detection performance is enhanced by combining the original model prediction with the output of the Hierarchical Ladder Network (HLN) via weighted probability fusion. To improve robustness under domain shift, we further introduce an Attention Affine Network (AAN) that adaptively refines the self-attention mechanism conditioned on the token information to better adapt to domain drift, thereby improving the classification performance of the model on datasets with domain shift. Additionally, a weighted entropy mechanism is employed to dynamically suppress the influence of low-confidence samples during adaptation. Experimental results on benchmark datasets show that our method significantly improves the performance on the most widely used classification datasets.

Yi Zhang, Chun-Wun Cheng, Junyi He et al.

We introduce SONO, a novel method leveraging Second-Order Neural Ordinary Differential Equations (Second-Order NODEs) to enhance cross-modal few-shot learning. By employing a simple yet effective architecture consisting of a Second-Order NODEs model paired with a cross-modal classifier, SONO addresses the significant challenge of overfitting, which is common in few-shot scenarios due to limited training examples. Our second-order approach can approximate a broader class of functions, enhancing the model's expressive power and feature generalization capabilities. We initialize our cross-modal classifier with text embeddings derived from class-relevant prompts, streamlining training efficiency by avoiding the need for frequent text encoder processing. Additionally, we utilize text-based image augmentation, exploiting CLIP's robust image-text correlation to enrich training data significantly. Extensive experiments across multiple datasets demonstrate that SONO outperforms existing state-of-the-art methods in few-shot learning performance.

Yushun Tang, Shuoshuo Chen, Zhehan Kan et al.

Fully test-time adaptation aims to adapt the network model based on sequential analysis of input samples during the inference stage to address the cross-domain performance degradation problem of deep neural networks. This work is based on the following interesting finding: in transformer-based image classification, the class token at the first transformer encoder layer can be learned to capture the domain-specific characteristics of target samples during test-time adaptation. This learned token, when combined with input image patch embeddings, is able to gradually remove the domain-specific information from the feature representations of input samples during the transformer encoding process, thereby significantly improving the test-time adaptation performance of the source model across different domains. We refer to this class token as visual conditioning token (VCT). To successfully learn the VCT, we propose a bi-level learning approach to capture the long-term variations of domain-specific characteristics while accommodating local variations of instance-specific characteristics. Experimental results on the benchmark datasets demonstrate that our proposed bi-level visual conditioning token learning method is able to achieve significantly improved test-time adaptation performance by up to 1.9%.

Yi Zhang, Ce Zhang, Zihan Liao et al.

Large-scale pre-trained Vision-Language Models (VLMs), such as CLIP and ALIGN, have introduced a new paradigm for learning transferable visual representations. Recently, there has been a surge of interest among researchers in developing lightweight fine-tuning techniques to adapt these models to downstream visual tasks. We recognize that current state-of-the-art fine-tuning methods, such as Tip-Adapter, simply consider the covariance between the query image feature and features of support few-shot training samples, which only captures linear relations and potentially instigates a deceptive perception of independence. To address this issue, in this work, we innovatively introduce Brownian Distance Covariance (BDC) to the field of vision-language reasoning. The BDC metric can model all possible relations, providing a robust metric for measuring feature dependence. Based on this, we present a novel method called BDC-Adapter, which integrates BDC prototype similarity reasoning and multi-modal reasoning network prediction to perform classification tasks. Our extensive experimental results show that the proposed BDC-Adapter can freely handle non-linear relations and fully characterize independence, outperforming the current state-of-the-art methods by large margins.

Yang Li, Shichao Kan, Zhihai He

Existing approaches for unsupervised metric learning focus on exploring self-supervision information within the input image itself. We observe that, when analyzing images, human eyes often compare images against each other instead of examining images individually. In addition, they often pay attention to certain keypoints, image regions, or objects which are discriminative between image classes but highly consistent within classes. Even if the image is being transformed, the attention pattern will be consistent. Motivated by this observation, we develop a new approach to unsupervised deep metric learning where the network is learned based on self-supervision information across images instead of within one single image. To characterize the consistent pattern of human attention during image comparisons, we introduce the idea of transformed attention consistency. It assumes that visually similar images, even undergoing different image transforms, should share the same consistent visual attention map. This consistency leads to a pairwise self-supervision loss, allowing us to learn a Siamese deep neural network to encode and compare images against their transformed or matched pairs. To further enhance the inter-class discriminative power of the feature generated by this network, we adapt the concept of triplet loss from supervised metric learning to our unsupervised case and introduce the contrastive clustering loss. Our extensive experimental results on benchmark datasets demonstrate that our proposed method outperforms current state-of-the-art methods for unsupervised metric learning by a large margin.

Jianhe Yuan, Zhihai He

Effective defense of deep neural networks against adversarial attacks remains a challenging problem, especially under powerful white-box attacks. In this paper, we develop a new method called ensemble generative cleaning with feedback loops (EGC-FL) for effective defense of deep neural networks. The proposed EGC-FL method is based on two central ideas. First, we introduce a transformed deadzone layer into the defense network, which consists of an orthonormal transform and a deadzone-based activation function, to destroy the sophisticated noise pattern of adversarial attacks. Second, by constructing a generative cleaning network with a feedback loop, we are able to generate an ensemble of diverse estimations of the original clean image. We then learn a network to fuse this set of diverse estimations together to restore the original image. Our extensive experimental results demonstrate that our approach improves the state-of-art by large margins in both white-box and black-box attacks. It significantly improves the classification accuracy for white-box PGD attacks upon the second best method by more than 29% on the SVHN dataset and more than 39% on the challenging CIFAR-10 dataset.

Guanghan Ning, Zhi Zhang, Xiaobo Ren et al.

In this work, we propose a joint audio-video fingerprint Automatic Content Recognition (ACR) technology for media retrieval. The problem is focused on how to balance the query accuracy and the size of fingerprint, and how to allocate the bits of the fingerprint to video frames and audio frames to achieve the best query accuracy. By constructing a novel concept called Coverage, which is highly correlated to the query accuracy, we are able to form a rate-coverage model to translate the original problem into an optimization problem that can be resolved by dynamic programming. To the best of our knowledge, this is the first work that uses joint audio-video fingerprint ACR technology for media retrieval with a theoretical problem formulation. Experimental results indicate that compared to reference algorithms, the proposed method has up to 25% query accuracy improvement while using 60% overall bit-rates, and 25% bit-rate reduction while achieving 85% accuracy, and it significantly outperforms the solution with single audio or video source fingerprint.