2022-10-19 更新
VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training
Authors:Zhan Tong, Yibing Song, Jue Wang, Limin Wang
Pre-training video transformers on extra large-scale datasets is generally required to achieve premier performance on relatively small datasets. In this paper, we show that video masked autoencoders (VideoMAE) are data-efficient learners for self-supervised video pre-training (SSVP). We are inspired by the recent ImageMAE and propose customized video tube masking with an extremely high ratio. This simple design makes video reconstruction a more challenging self-supervision task, thus encouraging extracting more effective video representations during this pre-training process. We obtain three important findings on SSVP: (1) An extremely high proportion of masking ratio (i.e., 90% to 95%) still yields favorable performance of VideoMAE. The temporally redundant video content enables a higher masking ratio than that of images. (2) VideoMAE achieves impressive results on very small datasets (i.e., around 3k-4k videos) without using any extra data. (3) VideoMAE shows that data quality is more important than data quantity for SSVP. Domain shift between pre-training and target datasets is an important issue. Notably, our VideoMAE with the vanilla ViT can achieve 87.4% on Kinetics-400, 75.4% on Something-Something V2, 91.3% on UCF101, and 62.6% on HMDB51, without using any extra data. Code is available at https://github.com/MCG-NJU/VideoMAE.
PDF NeurIPS 2022 camera-ready version
点此查看论文截图
Vision Transformer Visualization: What Neurons Tell and How Neurons Behave?
Authors:Van-Anh Nguyen, Khanh Pham Dinh, Long Tung Vuong, Thanh-Toan Do, Quan Hung Tran, Dinh Phung, Trung Le
Recently vision transformers (ViT) have been applied successfully for various tasks in computer vision. However, important questions such as why they work or how they behave still remain largely unknown. In this paper, we propose an effective visualization technique, to assist us in exposing the information carried in neurons and feature embeddings across the ViT’s layers. Our approach departs from the computational process of ViTs with a focus on visualizing the local and global information in input images and the latent feature embeddings at multiple levels. Visualizations at the input and embeddings at level 0 reveal interesting findings such as providing support as to why ViTs are rather generally robust to image occlusions and patch shuffling; or unlike CNNs, level 0 embeddings already carry rich semantic details. Next, we develop a rigorous framework to perform effective visualizations across layers, exposing the effects of ViTs filters and grouping/clustering behaviors to object patches. Finally, we provide comprehensive experiments on real datasets to qualitatively and quantitatively demonstrate the merit of our proposed methods as well as our findings. https://github.com/byM1902/ViT_visualization
PDF The first two authors contributed equally to this work. Our code is available at https://github.com/byM1902/ViT_visualization
点此查看论文截图
UVCGAN: UNet Vision Transformer cycle-consistent GAN for unpaired image-to-image translation
Authors:Dmitrii Torbunov, Yi Huang, Haiwang Yu, Jin Huang, Shinjae Yoo, Meifeng Lin, Brett Viren, Yihui Ren
Unpaired image-to-image translation has broad applications in art, design, and scientific simulations. One early breakthrough was CycleGAN that emphasizes one-to-one mappings between two unpaired image domains via generative-adversarial networks (GAN) coupled with the cycle-consistency constraint, while more recent works promote one-to-many mapping to boost diversity of the translated images. Motivated by scientific simulation and one-to-one needs, this work revisits the classic CycleGAN framework and boosts its performance to outperform more contemporary models without relaxing the cycle-consistency constraint. To achieve this, we equip the generator with a Vision Transformer (ViT) and employ necessary training and regularization techniques. Compared to previous best-performing models, our model performs better and retains a strong correlation between the original and translated image. An accompanying ablation study shows that both the gradient penalty and self-supervised pre-training are crucial to the improvement. To promote reproducibility and open science, the source code, hyperparameter configurations, and pre-trained model are available at https://github.com/LS4GAN/uvcgan.
PDF Accepted by WACV2023, contains 5 pages, 2 figures, 2 tables
点此查看论文截图
Sequence and Circle: Exploring the Relationship Between Patches
Authors:Zhengyang Yu, Jochen Triesch
The vision transformer (ViT) has achieved state-of-the-art results in various vision tasks. It utilizes a learnable position embedding (PE) mechanism to encode the location of each image patch. However, it is presently unclear if this learnable PE is really necessary and what its benefits are. This paper explores two alternative ways of encoding the location of individual patches that exploit prior knowledge about their spatial arrangement. One is called the sequence relationship embedding (SRE), and the other is called the circle relationship embedding(CRE). Among them, the SRE considers all patches to be in order, and adjacent patches have the same interval distance. The CRE considers the central patch as the center of the circle and measures the distance of the remaining patches from the center based on the four neighborhoods principle. Multiple concentric circles with different radii combine different patches. Finally, we implemented these two relations on three classic ViTs and tested them on four popular datasets. Experiments show that SRE and CRE can replace PE to reduce the random learnable parameters while achieving the same performance. Combining SRE or CRE with PE gets better performance than only using PE.
PDF 7 pages, 1 figure
点此查看论文截图
Token Merging: Your ViT But Faster
Authors:Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, Judy Hoffman
We introduce Token Merging (ToMe), a simple method to increase the throughput of existing ViT models without needing to train. ToMe gradually combines similar tokens in a transformer using a general and light-weight matching algorithm that is as fast as pruning while being more accurate. Off-the-shelf, ToMe can 2x the throughput of state-of-the-art ViT-L @ 512 and ViT-H @ 518 models on images and 2.2x the throughput of ViT-L on video with only a 0.2-0.3% accuracy drop in each case. ToMe can also easily be applied during training, improving in practice training speed up to 2x for MAE fine-tuning on video. Training with ToMe further minimizes accuracy drop, leading to 2x the throughput of ViT-B on audio for only a 0.4% mAP drop. Qualitatively, we find that ToMe merges object parts into one token, even over multiple frames of video. Overall, ToMe’s accuracy and speed are competitive with state-of-the-art on images, video, and audio.
PDF Preprint. Code will be available here: https://github.com/facebookresearch/ToMe
点此查看论文截图