2023-11-05 更新
SYRAC: Synthesize, Rank, and Count
Authors:Adriano D’Alessandro, Ali Mahdavi-Amiri, Ghassan Hamarneh
Crowd counting is a critical task in computer vision, with several important applications. However, existing counting methods rely on labor-intensive density map annotations, necessitating the manual localization of each individual pedestrian. While recent efforts have attempted to alleviate the annotation burden through weakly or semi-supervised learning, these approaches fall short of significantly reducing the workload. We propose a novel approach to eliminate the annotation burden by leveraging latent diffusion models to generate synthetic data. However, these models struggle to reliably understand object quantities, leading to noisy annotations when prompted to produce images with a specific quantity of objects. To address this, we use latent diffusion models to create two types of synthetic data: one by removing pedestrians from real images, which generates ranked image pairs with a weak but reliable object quantity signal, and the other by generating synthetic images with a predetermined number of objects, offering a strong but noisy counting signal. Our method utilizes the ranking image pairs for pre-training and then fits a linear layer to the noisy synthetic images using these crowd quantity features. We report state-of-the-art results for unsupervised crowd counting.
PDF
点此查看论文截图
LangNav: Language as a Perceptual Representation for Navigation
Authors:Bowen Pan, Rameswar Panda, SouYoung Jin, Rogerio Feris, Aude Oliva, Phillip Isola, Yoon Kim
We explore the use of language as a perceptual representation for vision-and-language navigation. Our approach uses off-the-shelf vision systems (for image captioning and object detection) to convert an agent’s egocentric panoramic view at each time step into natural language descriptions. We then finetune a pretrained language model to select an action, based on the current view and the trajectory history, that would best fulfill the navigation instructions. In contrast to the standard setup which adapts a pretrained language model to work directly with continuous visual features from pretrained vision models, our approach instead uses (discrete) language as the perceptual representation. We explore two use cases of our language-based navigation (LangNav) approach on the R2R vision-and-language navigation benchmark: generating synthetic trajectories from a prompted large language model (GPT-4) with which to finetune a smaller language model; and sim-to-real transfer where we transfer a policy learned on a simulated environment (ALFRED) to a real-world environment (R2R). Our approach is found to improve upon strong baselines that rely on visual features in settings where only a few gold trajectories (10-100) are available, demonstrating the potential of using language as a perceptual representation for navigation tasks.
PDF
点此查看论文截图
Heart Disease Detection using Vision-Based Transformer Models from ECG Images
Authors:Zeynep Hilal Kilimci, Mustafa Yalcin, Ayhan Kucukmanisa, Amit Kumar Mishra
Heart disease, also known as cardiovascular disease, is a prevalent and critical medical condition characterized by the impairment of the heart and blood vessels, leading to various complications such as coronary artery disease, heart failure, and myocardial infarction. The timely and accurate detection of heart disease is of paramount importance in clinical practice. Early identification of individuals at risk enables proactive interventions, preventive measures, and personalized treatment strategies to mitigate the progression of the disease and reduce adverse outcomes. In recent years, the field of heart disease detection has witnessed notable advancements due to the integration of sophisticated technologies and computational approaches. These include machine learning algorithms, data mining techniques, and predictive modeling frameworks that leverage vast amounts of clinical and physiological data to improve diagnostic accuracy and risk stratification. In this work, we propose to detect heart disease from ECG images using cutting-edge technologies, namely vision transformer models. These models are Google-Vit, Microsoft-Beit, and Swin-Tiny. To the best of our knowledge, this is the initial endeavor concentrating on the detection of heart diseases through image-based ECG data by employing cuttingedge technologies namely, transformer models. To demonstrate the contribution of the proposed framework, the performance of vision transformer models are compared with state-of-the-art studies. Experiment results show that the proposed framework exhibits remarkable classification results.
PDF
点此查看论文截图
Minimalist and High-Performance Semantic Segmentation with Plain Vision Transformers
Authors:Yuanduo Hong, Jue Wang, Weichao Sun, Huihui Pan
In the wake of Masked Image Modeling (MIM), a diverse range of plain, non-hierarchical Vision Transformer (ViT) models have been pre-trained with extensive datasets, offering new paradigms and significant potential for semantic segmentation. Current state-of-the-art systems incorporate numerous inductive biases and employ cumbersome decoders. Building upon the original motivations of plain ViTs, which are simplicity and generality, we explore high-performance `minimalist’ systems to this end. Our primary purpose is to provide simple and efficient baselines for practical semantic segmentation with plain ViTs. Specifically, we first explore the feasibility and methodology for achieving high-performance semantic segmentation using the last feature map. As a result, we introduce the PlainSeg, a model comprising only three 3$\times$3 convolutions in addition to the transformer layers (either encoder or decoder). In this process, we offer insights into two underlying principles: (i) high-resolution features are crucial to high performance in spite of employing simple up-sampling techniques and (ii) the slim transformer decoder requires a much larger learning rate than the wide transformer decoder. On this basis, we further present the PlainSeg-Hier, which allows for the utilization of hierarchical features. Extensive experiments on four popular benchmarks demonstrate the high performance and efficiency of our methods. They can also serve as powerful tools for assessing the transfer ability of base models in semantic segmentation. Code is available at \url{https://github.com/ydhongHIT/PlainSeg}.
PDF
点此查看论文截图
Unsupervised Object Localization in the Era of Self-Supervised ViTs: A Survey
Authors:Oriane Siméoni, Éloi Zablocki, Spyros Gidaris, Gilles Puy, Patrick Pérez
The recent enthusiasm for open-world vision systems show the high interest of the community to perform perception tasks outside of the closed-vocabulary benchmark setups which have been so popular until now. Being able to discover objects in images/videos without knowing in advance what objects populate the dataset is an exciting prospect. But how to find objects without knowing anything about them? Recent works show that it is possible to perform class-agnostic unsupervised object localization by exploiting self-supervised pre-trained features. We propose here a survey of unsupervised object localization methods that discover objects in images without requiring any manual annotation in the era of self-supervised ViTs. We gather links of discussed methods in the repository https://github.com/valeoai/Awesome-Unsupervised-Object-Localization.
PDF
点此查看论文截图
CXR-CLIP: Toward Large Scale Chest X-ray Language-Image Pre-training
Authors:Kihyun You, Jawook Gu, Jiyeon Ham, Beomhee Park, Jiho Kim, Eun Kyoung Hong, Woonhyunk Baek, Byungseok Roh
A large-scale image-text pair dataset has greatly contributed to the development of vision-language pre-training (VLP) models, which enable zero-shot or few-shot classification without costly annotation. However, in the medical domain, the scarcity of data remains a significant challenge for developing a powerful VLP model. In this paper, we tackle the lack of image-text data in chest X-ray by expanding image-label pair as image-text pair via general prompt and utilizing multiple images and multiple sections in a radiologic report. We also design two contrastive losses, named ICL and TCL, for learning study-level characteristics of medical images and reports, respectively. Our model outperforms the state-of-the-art models trained under the same conditions. Also, enlarged dataset improve the discriminative power of our pre-trained model for classification, while sacrificing marginal retrieval performance. Code is available at https://github.com/kakaobrain/cxr-clip.
PDF Accepted by MICCAI 2023
点此查看论文截图
Exploring Driving Behavior for Autonomous Vehicles Based on Gramian Angular Field Vision Transformer
Authors:Junwei You, Ying Chen, Zhuoyu Jiang, Zhangchi Liu, Zilin Huang, Yifeng Ding, Bin Ran
Effective classification of autonomous vehicle (AV) driving behavior emerges as a critical area for diagnosing AV operation faults, enhancing autonomous driving algorithms, and reducing accident rates. This paper presents the Gramian Angular Field Vision Transformer (GAF-ViT) model, designed to analyze AV driving behavior. The proposed GAF-ViT model consists of three key components: GAF Transformer Module, Channel Attention Module, and Multi-Channel ViT Module. These modules collectively convert representative sequences of multivariate behavior into multi-channel images and employ image recognition techniques for behavior classification. A channel attention mechanism is applied to multi-channel images to discern the impact of various driving behavior features. Experimental evaluation on the Waymo Open Dataset of trajectories demonstrates that the proposed model achieves state-of-the-art performance. Furthermore, an ablation study effectively substantiates the efficacy of individual modules within the model.
PDF
点此查看论文截图
Generalizing to Unseen Domains in Diabetic Retinopathy Classification
Authors:Chamuditha Jayanga Galappaththige, Gayal Kuruppu, Muhammad Haris Khan
Diabetic retinopathy (DR) is caused by long-standing diabetes and is among the fifth leading cause for visual impairments. The process of early diagnosis and treatments could be helpful in curing the disease, however, the detection procedure is rather challenging and mostly tedious. Therefore, automated diabetic retinopathy classification using deep learning techniques has gained interest in the medical imaging community. Akin to several other real-world applications of deep learning, the typical assumption of i.i.d data is also violated in DR classification that relies on deep learning. Therefore, developing DR classification methods robust to unseen distributions is of great value. In this paper, we study the problem of generalizing a model to unseen distributions or domains (a.k.a domain generalization) in DR classification. To this end, we propose a simple and effective domain generalization (DG) approach that achieves self-distillation in vision transformers (ViT) via a novel prediction softening mechanism. This prediction softening is an adaptive convex combination one-hot labels with the model’s own knowledge. We perform extensive experiments on challenging open-source DR classification datasets under both multi-source and single-source DG settings with three different ViT backbones to establish the efficacy and applicability of our approach against competing methods. For the first time, we report the performance of several state-of-the-art DG methods on open-source DR classification datasets after conducting thorough experiments. Finally, our method is also capable of delivering improved calibration performance than other methods, showing its suitability for safety-critical applications, including healthcare. We hope that our contributions would investigate more DG research across the medical imaging community.
PDF Accepted at WACV 2024
点此查看论文截图
AntifakePrompt: Prompt-Tuned Vision-Language Models are Fake Image Detectors
Authors:You-Ming Chang, Chen Yeh, Wei-Chen Chiu, Ning Yu
Deep generative models can create remarkably photorealistic fake images while raising concerns about misinformation and copyright infringement, known as deepfake threats. Deepfake detection technique is developed to distinguish between real and fake images, where the existing methods typically learn classifiers in the image domain or various feature domains. However, the generalizability of deepfake detection against emerging and more advanced generative models remains challenging. In this paper, being inspired by the zero-shot advantages of Vision-Language Models (VLMs), we propose a novel approach using VLMs (e.g. InstructBLIP) and prompt tuning techniques to improve the deepfake detection accuracy over unseen data. We formulate deepfake detection as a visual question answering problem, and tune soft prompts for InstructBLIP to answer the real/fake information of a query image. We conduct full-spectrum experiments on datasets from 3 held-in and 13 held-out generative models, covering modern text-to-image generation, image editing and image attacks. Results demonstrate that (1) the deepfake detection accuracy can be significantly and consistently improved (from 58.8% to 91.31%, in average accuracy over unseen data) using pretrained vision-language models with prompt tuning; (2) our superior performance is at less cost of trainable parameters, resulting in an effective and efficient solution for deepfake detection. Code and models can be found at https://github.com/nctu-eva-lab/AntifakePrompt.
PDF
点此查看论文截图
Sliceformer: Make Multi-head Attention as Simple as Sorting in Discriminative Tasks
Authors:Shen Yuan, Hongteng Xu
As one of the most popular neural network modules, Transformer plays a central role in many fundamental deep learning models, e.g., the ViT in computer vision and the BERT and GPT in natural language processing. The effectiveness of the Transformer is often attributed to its multi-head attention (MHA) mechanism. In this study, we discuss the limitations of MHA, including the high computational complexity due to its query-key-value'' architecture and the numerical issue caused by its softmax operation. Considering the above problems and the recent development tendency of the attention layer, we propose an effective and efficient surrogate of the Transformer, called Sliceformer. Our Sliceformer replaces the classic MHA mechanism with an extremely simpleslicing-sorting’’ operation, i.e., projecting inputs linearly to a latent space and sorting them along different feature dimensions (or equivalently, called channels). For each feature dimension, the sorting operation implicitly generates an implicit attention map with sparse, full-rank, and doubly-stochastic structures. We consider different implementations of the slicing-sorting operation and analyze their impacts on the Sliceformer. We test the Sliceformer in the Long-Range Arena benchmark, image classification, text classification, and molecular property prediction, demonstrating its advantage in computational complexity and universal effectiveness in discriminative tasks. Our Sliceformer achieves comparable or better performance with lower memory cost and faster speed than the Transformer and its variants. Moreover, the experimental results reveal that applying our Sliceformer can empirically suppress the risk of mode collapse when representing data. The code is available at \url{https://github.com/SDS-Lab/sliceformer}.
PDF
点此查看论文截图
LipSim: A Provably Robust Perceptual Similarity Metric
Authors:Sara Ghazanfari, Alexandre Araujo, Prashanth Krishnamurthy, Farshad Khorrami, Siddharth Garg
Recent years have seen growing interest in developing and applying perceptual similarity metrics. Research has shown the superiority of perceptual metrics over pixel-wise metrics in aligning with human perception and serving as a proxy for the human visual system. On the other hand, as perceptual metrics rely on neural networks, there is a growing concern regarding their resilience, given the established vulnerability of neural networks to adversarial attacks. It is indeed logical to infer that perceptual metrics may inherit both the strengths and shortcomings of neural networks. In this work, we demonstrate the vulnerability of state-of-the-art perceptual similarity metrics based on an ensemble of ViT-based feature extractors to adversarial attacks. We then propose a framework to train a robust perceptual similarity metric called LipSim (Lipschitz Similarity Metric) with provable guarantees. By leveraging 1-Lipschitz neural networks as the backbone, LipSim provides guarded areas around each data point and certificates for all perturbations within an $\ell_2$ ball. Finally, a comprehensive set of experiments shows the performance of LipSim in terms of natural and certified scores and on the image retrieval application. The code is available at https://github.com/SaraGhazanfari/LipSim.
PDF
点此查看论文截图
One-shot Localization and Segmentation of Medical Images with Foundation Models
Authors:Deepa Anand, Gurunath Reddy M, Vanika Singhal, Dattesh D. Shanbhag, Shriram KS, Uday Patil, Chitresh Bhushan, Kavitha Manickam, Dawei Gui, Rakesh Mullick, Avinash Gopal, Parminder Bhatia, Taha Kass-Hout
Recent advances in Vision Transformers (ViT) and Stable Diffusion (SD) models with their ability to capture rich semantic features of the image have been used for image correspondence tasks on natural images. In this paper, we examine the ability of a variety of pre-trained ViT (DINO, DINOv2, SAM, CLIP) and SD models, trained exclusively on natural images, for solving the correspondence problems on medical images. While many works have made a case for in-domain training, we show that the models trained on natural images can offer good performance on medical images across different modalities (CT,MR,Ultrasound) sourced from various manufacturers, over multiple anatomical regions (brain, thorax, abdomen, extremities), and on wide variety of tasks. Further, we leverage the correspondence with respect to a template image to prompt a Segment Anything (SAM) model to arrive at single shot segmentation, achieving dice range of 62%-90% across tasks, using just one image as reference. We also show that our single-shot method outperforms the recently proposed few-shot segmentation method - UniverSeg (Dice range 47%-80%) on most of the semantic segmentation tasks(six out of seven) across medical imaging modalities.
PDF Accepted at NeurIPS 2023 R0-FoMo Workshop
点此查看论文截图
Local-Global Self-Supervised Visual Representation Learning
Authors:Ali Javidani, Mohammad Amin Sadeghi, Babak Nadjar Araabi
Self-supervised representation learning methods mainly focus on image-level instance discrimination. This study explores the potential benefits of incorporating patch-level discrimination into existing methods to enhance the quality of learned representations by simultaneously looking at local and global visual features. Towards this idea, we present a straightforward yet effective patch-matching algorithm that can find the corresponding patches across the augmented views of an image. The augmented views are subsequently fed into a self-supervised learning framework employing Vision Transformer (ViT) as its backbone. The result is the generation of both image-level and patch-level representations. Leveraging the proposed patch-matching algorithm, the model minimizes the representation distance between not only the CLS tokens but also the corresponding patches. As a result, the model gains a more comprehensive understanding of both the entirety of the image as well as its finer details. We pretrain the proposed method on small, medium, and large-scale datasets. It is shown that our approach could outperform state-of-the-art image-level representation learning methods on both image classification and downstream tasks. Keywords: Self-Supervised Learning; Visual Representations; Local-Global Representation Learning; Patch-Wise Representation Learning; Vision Transformer (ViT)
PDF
点此查看论文截图
AnomalyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection
Authors:Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, Jiming Chen
Zero-shot anomaly detection (ZSAD) requires detection models trained using auxiliary data to detect anomalies without any training sample in a target dataset. It is a crucial task when training data is not accessible due to various concerns, \eg, data privacy, yet it is challenging since the models need to generalize to anomalies across different domains where the appearance of foreground objects, abnormal regions, and background features, such as defects/tumors on different products/organs, can vary significantly. Recently large pre-trained vision-language models (VLMs), such as CLIP, have demonstrated strong zero-shot recognition ability in various vision tasks, including anomaly detection. However, their ZSAD performance is weak since the VLMs focus more on modeling the class semantics of the foreground objects rather than the abnormality/normality in the images. In this paper we introduce a novel approach, namely AnomalyCLIP, to adapt CLIP for accurate ZSAD across different domains. The key insight of AnomalyCLIP is to learn object-agnostic text prompts that capture generic normality and abnormality in an image regardless of its foreground objects. This allows our model to focus on the abnormal image regions rather than the object semantics, enabling generalized normality and abnormality recognition on diverse types of objects. Large-scale experiments on 17 real-world anomaly detection datasets show that AnomalyCLIP achieves superior zero-shot performance of detecting and segmenting anomalies in datasets of highly diverse class semantics from various defect inspection and medical imaging domains. Code will be made available at https://github.com/zqhang/AnomalyCLIP.
PDF
点此查看论文截图
BirdSAT: Cross-View Contrastive Masked Autoencoders for Bird Species Classification and Mapping
Authors:Srikumar Sastry, Subash Khanal, Aayush Dhakal, Di Huang, Nathan Jacobs
We propose a metadata-aware self-supervised learning~(SSL)~framework useful for fine-grained classification and ecological mapping of bird species around the world. Our framework unifies two SSL strategies: Contrastive Learning~(CL) and Masked Image Modeling~(MIM), while also enriching the embedding space with metadata available with ground-level imagery of birds. We separately train uni-modal and cross-modal ViT on a novel cross-view global bird species dataset containing ground-level imagery, metadata (location, time), and corresponding satellite imagery. We demonstrate that our models learn fine-grained and geographically conditioned features of birds, by evaluating on two downstream tasks: fine-grained visual classification~(FGVC) and cross-modal retrieval. Pre-trained models learned using our framework achieve SotA performance on FGVC of iNAT-2021 birds and in transfer learning settings for CUB-200-2011 and NABirds datasets. Moreover, the impressive cross-modal retrieval performance of our model enables the creation of species distribution maps across any geographic region. The dataset and source code will be released at https://github.com/mvrl/BirdSAT}.
PDF Accepted at WACV 2024
点此查看论文截图
Class Incremental Learning with Pre-trained Vision-Language Models
Authors:Xialei Liu, Xusheng Cao, Haori Lu, Jia-wen Xiao, Andrew D. Bagdanov, Ming-Ming Cheng
With the advent of large-scale pre-trained models, interest in adapting and exploiting them for continual learning scenarios has grown. In this paper, we propose an approach to exploiting pre-trained vision-language models (e.g. CLIP) that enables further adaptation instead of only using zero-shot learning of new tasks. We augment a pre-trained CLIP model with additional layers after the Image Encoder or before the Text Encoder. We investigate three different strategies: a Linear Adapter, a Self-attention Adapter, each operating on the image embedding, and Prompt Tuning which instead modifies prompts input to the CLIP text encoder. We also propose a method for parameter retention in the adapter layers that uses a measure of parameter importance to better maintain stability and plasticity during incremental learning. Our experiments demonstrate that the simplest solution — a single Linear Adapter layer with parameter retention — produces the best results. Experiments on several conventional benchmarks consistently show a significant margin of improvement over the current state-of-the-art.
PDF
点此查看论文截图
HAP: Structure-Aware Masked Image Modeling for Human-Centric Perception
Authors:Junkun Yuan, Xinyu Zhang, Hao Zhou, Jian Wang, Zhongwei Qiu, Zhiyin Shao, Shaofeng Zhang, Sifan Long, Kun Kuang, Kun Yao, Junyu Han, Errui Ding, Lanfen Lin, Fei Wu, Jingdong Wang
Model pre-training is essential in human-centric perception. In this paper, we first introduce masked image modeling (MIM) as a pre-training approach for this task. Upon revisiting the MIM training strategy, we reveal that human structure priors offer significant potential. Motivated by this insight, we further incorporate an intuitive human structure prior - human parts - into pre-training. Specifically, we employ this prior to guide the mask sampling process. Image patches, corresponding to human part regions, have high priority to be masked out. This encourages the model to concentrate more on body structure information during pre-training, yielding substantial benefits across a range of human-centric perception tasks. To further capture human characteristics, we propose a structure-invariant alignment loss that enforces different masked views, guided by the human part prior, to be closely aligned for the same image. We term the entire method as HAP. HAP simply uses a plain ViT as the encoder yet establishes new state-of-the-art performance on 11 human-centric benchmarks, and on-par result on one dataset. For example, HAP achieves 78.1% mAP on MSMT17 for person re-identification, 86.54% mA on PA-100K for pedestrian attribute recognition, 78.2% AP on MS COCO for 2D pose estimation, and 56.0 PA-MPJPE on 3DPW for 3D pose and shape estimation.
PDF Accepted by NeurIPS 2023
点此查看论文截图
Improving Robustness for Vision Transformer with a Simple Dynamic Scanning Augmentation
Authors:Shashank Kotyan, Danilo Vasconcellos Vargas
Vision Transformer (ViT) has demonstrated promising performance in computer vision tasks, comparable to state-of-the-art neural networks. Yet, this new type of deep neural network architecture is vulnerable to adversarial attacks limiting its capabilities in terms of robustness. This article presents a novel contribution aimed at further improving the accuracy and robustness of ViT, particularly in the face of adversarial attacks. We propose an augmentation technique called `Dynamic Scanning Augmentation’ that leverages dynamic input sequences to adaptively focus on different patches, thereby maintaining performance and robustness. Our detailed investigations reveal that this adaptability to the input sequence induces significant changes in the attention mechanism of ViT, even for the same image. We introduce four variations of Dynamic Scanning Augmentation, outperforming ViT in terms of both robustness to adversarial attacks and accuracy against natural images, with one variant showing comparable results. By integrating our augmentation technique, we observe a substantial increase in ViT’s robustness, improving it from $17\%$ to $92\%$ measured across different types of adversarial attacks. These findings, together with other comprehensive tests, indicate that Dynamic Scanning Augmentation enhances accuracy and robustness by promoting a more adaptive type of attention. In conclusion, this work contributes to the ongoing research on Vision Transformers by introducing Dynamic Scanning Augmentation as a technique for improving the accuracy and robustness of ViT. The observed results highlight the potential of this approach in advancing computer vision tasks and merit further exploration in future studies.
PDF Accepted in Neurocomputing
点此查看论文截图
Learning to Adapt CLIP for Few-Shot Monocular Depth Estimation
Authors:Xueting Hu, Ce Zhang, Yi Zhang, Bowen Hai, Ke Yu, Zhihai He
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.
PDF Accepted by WACV 2024
点此查看论文截图
AiluRus: A Scalable ViT Framework for Dense Prediction
Authors:Jin Li, Yaoming Wang, Xiaopeng Zhang, Bowen Shi, Dongsheng Jiang, Chenglin Li, Wenrui Dai, Hongkai Xiong, Qi Tian
Vision transformers (ViTs) have emerged as a prevalent architecture for vision tasks owing to their impressive performance. However, when it comes to handling long token sequences, especially in dense prediction tasks that require high-resolution input, the complexity of ViTs increases significantly. Notably, dense prediction tasks, such as semantic segmentation or object detection, emphasize more on the contours or shapes of objects, while the texture inside objects is less informative. Motivated by this observation, we propose to apply adaptive resolution for different regions in the image according to their importance. Specifically, at the intermediate layer of the ViT, we utilize a spatial-aware density-based clustering algorithm to select representative tokens from the token sequence. Once the representative tokens are determined, we proceed to merge other tokens into their closest representative token. Consequently, semantic similar tokens are merged together to form low-resolution regions, while semantic irrelevant tokens are preserved independently as high-resolution regions. This strategy effectively reduces the number of tokens, allowing subsequent layers to handle a reduced token sequence and achieve acceleration. We evaluate our proposed method on three different datasets and observe promising performance. For example, the “Segmenter ViT-L” model can be accelerated by 48% FPS without fine-tuning, while maintaining the performance. Additionally, our method can be applied to accelerate fine-tuning as well. Experimental results demonstrate that we can save 52% training time while accelerating 2.46 times FPS with only a 0.09% performance drop. The code is available at https://github.com/caddyless/ailurus/tree/main.
PDF Accepted by NeurIPS 2023
点此查看论文截图