2022-04-01 更新
Visual Prompting: Modifying Pixel Space to Adapt Pre-trained Models
Authors:Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, Phillip Isola
Prompting has recently become a popular paradigm for adapting language models to downstream tasks. Rather than fine-tuning model parameters or adding task-specific heads, this approach steers a model to perform a new task simply by adding a text prompt to the model’s inputs. In this paper, we explore the question: can we create prompts with pixels instead? In other words, can pre-trained vision models be adapted to a new task solely by adding pixels to their inputs? We introduce visual prompting, which learns a task-specific image perturbation such that a frozen pre-trained model prompted with this perturbation performs a new task. We discover that changing only a few pixels is enough to adapt models to new tasks and datasets, and performs on par with linear probing, the current de facto approach to lightweight adaptation. The surprising effectiveness of visual prompting provides a new perspective on how to adapt pre-trained models in vision, and opens up the possibility of adapting models solely through their inputs, which, unlike model parameters or outputs, are typically under an end-user’s control. Code is available at http://hjbahng.github.io/visual_prompting .
PDF 17 pages, 10 figures
论文截图
RegionViT: Regional-to-Local Attention for Vision Transformers
Authors:Chun-Fu Chen, Rameswar Panda, Quanfu Fan
Vision transformer (ViT) has recently shown its strong capability in achieving comparable results to convolutional neural networks (CNNs) on image classification. However, vanilla ViT simply inherits the same architecture from the natural language processing directly, which is often not optimized for vision applications. Motivated by this, in this paper, we propose a new architecture that adopts the pyramid structure and employ a novel regional-to-local attention rather than global self-attention in vision transformers. More specifically, our model first generates regional tokens and local tokens from an image with different patch sizes, where each regional token is associated with a set of local tokens based on the spatial location. The regional-to-local attention includes two steps: first, the regional self-attention extract global information among all regional tokens and then the local self-attention exchanges the information among one regional token and the associated local tokens via self-attention. Therefore, even though local self-attention confines the scope in a local region but it can still receive global information. Extensive experiments on four vision tasks, including image classification, object and keypoint detection, semantics segmentation and action recognition, show that our approach outperforms or is on par with state-of-the-art ViT variants including many concurrent works. Our source codes and models are available at https://github.com/ibm/regionvit.
PDF add more results and link to codes and models. https://github.com/ibm/regionvit, formatted with ICLR style
论文截图
Multimodal Fusion Transformer for Remote Sensing Image Classification
Authors:Swalpa Kumar Roy, Ankur Deria, Danfeng Hong, Behnood Rasti, Antonio Plaza, Jocelyn Chanussot
Vision transformer (ViT) has been trending in image classification tasks due to its promising performance when compared to convolutional neural networks (CNNs). As a result, many researchers have tried to incorporate ViT models in hyperspectral image (HSI) classification tasks, but without achieving satisfactory performance. To this paper, we introduce a new multimodal fusion transformer (MFT) network for HSI land-cover classification, which utilizes other sources of multimodal data in addition to HSI. Instead of using conventional feature fusion techniques, other multimodal data are used as an external classification (CLS) token in the transformer encoder, which helps achieving better generalization. ViT and other similar transformer models use a randomly initialized external classification token {and fail to generalize well}. However, the use of a feature embedding derived from other sources of multimodal data, such as light detection and ranging (LiDAR), offers the potential to improve those models by means of a CLS. The concept of tokenization is used in our work to generate CLS and HSI patch tokens, helping to learn key features in a reduced feature space. We also introduce a new attention mechanism for improving the exchange of information between HSI tokens and the CLS (e.g., LiDAR) token. Extensive experiments are carried out on widely used and benchmark datasets i.e., the University of Houston, Trento, University of Southern Mississippi Gulfpark (MUUFL), and Augsburg. In the results section, we compare the proposed MFT model with other state-of-the-art transformer models, classical CNN models, as well as conventional classifiers. The superior performance achieved by the proposed model is due to the use of multimodal information as external classification tokens.
PDF
论文截图