Zhan Tong

Chongjian Ge, Xiaohan Ding, Zhan Tong et al.

Vision Transformers (ViTs) have been shown to enhance visual recognition through modeling long-range dependencies with multi-head self-attention (MHSA), which is typically formulated as Query-Key-Value computation. However, the attention map generated from the Query and Key captures only token-to-token correlations at one single granularity. In this paper, we argue that self-attention should have a more comprehensive mechanism to capture correlations among tokens and groups (i.e., multiple adjacent tokens) for higher representational capacity. Thereby, we propose Group-Mix Attention (GMA) as an advanced replacement for traditional self-attention, which can simultaneously capture token-to-token, token-to-group, and group-to-group correlations with various group sizes. To this end, GMA splits the Query, Key, and Value into segments uniformly and performs different group aggregations to generate group proxies. The attention map is computed based on the mixtures of tokens and group proxies and used to re-combine the tokens and groups in Value. Based on GMA, we introduce a powerful backbone, namely GroupMixFormer, which achieves state-of-the-art performance in image classification, object detection, and semantic segmentation with fewer parameters than existing models. For instance, GroupMixFormer-L (with 70.3M parameters and 384^2 input) attains 86.2% Top-1 accuracy on ImageNet-1K without external data, while GroupMixFormer-B (with 45.8M parameters) attains 51.2% mIoU on ADE20K.

Ziteng Gao, Zhan Tong, Kevin Qinghong Lin et al.

The recently proposed SparseFormer architecture provides an alternative approach to visual understanding by utilizing a significantly lower number of visual tokens via adjusting RoIs, greatly reducing computational costs while still achieving promising performance. However, training SparseFormers from scratch is still expensive, and scaling up the number of parameters can be challenging. In this paper, we propose to bootstrap SparseFormers from ViT-based vision foundation models in a simple and efficient way. Since the majority of SparseFormer blocks are the standard transformer ones, we can inherit weights from large-scale pre-trained vision transformers and freeze them as much as possible. Therefore, we only need to train the SparseFormer-specific lightweight focusing transformer to adjust token RoIs and fine-tune a few early pre-trained blocks to align the final token representation. In such a way, we can bootstrap SparseFormer architectures from various large-scale pre-trained models (e.g., IN-21K pre-trained AugRegs or CLIPs) using a rather smaller amount of training samples (e.g., IN-1K) and without labels or captions within just a few hours. As a result, the bootstrapped unimodal SparseFormer (from AugReg-ViT-L/16-384) can reach 84.9% accuracy on IN-1K with only 49 tokens, and the multimodal SparseFormer from CLIPs also demonstrates notable zero-shot performance with highly reduced computational cost without seeing any caption during the bootstrapping procedure. In addition, CLIP-bootstrapped SparseFormers, which align the output space with language without seeing a word, can serve as efficient vision encoders in multimodal large language models. Code and models are available at https://github.com/showlab/sparseformer

Youwei Liang, Chongjian Ge, Zhan Tong et al.

Vision Transformers (ViTs) take all the image patches as tokens and construct multi-head self-attention (MHSA) among them. Complete leverage of these image tokens brings redundant computations since not all the tokens are attentive in MHSA. Examples include that tokens containing semantically meaningless or distractive image backgrounds do not positively contribute to the ViT predictions. In this work, we propose to reorganize image tokens during the feed-forward process of ViT models, which is integrated into ViT during training. For each forward inference, we identify the attentive image tokens between MHSA and FFN (i.e., feed-forward network) modules, which is guided by the corresponding class token attention. Then, we reorganize image tokens by preserving attentive image tokens and fusing inattentive ones to expedite subsequent MHSA and FFN computations. To this end, our method EViT improves ViTs from two perspectives. First, under the same amount of input image tokens, our method reduces MHSA and FFN computation for efficient inference. For instance, the inference speed of DeiT-S is increased by 50% while its recognition accuracy is decreased by only 0.3% for ImageNet classification. Second, by maintaining the same computational cost, our method empowers ViTs to take more image tokens as input for recognition accuracy improvement, where the image tokens are from higher resolution images. An example is that we improve the recognition accuracy of DeiT-S by 1% for ImageNet classification at the same computational cost of a vanilla DeiT-S. Meanwhile, our method does not introduce more parameters to ViTs. Experiments on the standard benchmarks show the effectiveness of our method. The code is available at https://github.com/youweiliang/evit

Limin Wang, Zhan Tong, Bin Ji et al.

Temporal modeling still remains challenging for action recognition in videos. To mitigate this issue, this paper presents a new video architecture, termed as Temporal Difference Network (TDN), with a focus on capturing multi-scale temporal information for efficient action recognition. The core of our TDN is to devise an efficient temporal module (TDM) by explicitly leveraging a temporal difference operator, and systematically assess its effect on short-term and long-term motion modeling. To fully capture temporal information over the entire video, our TDN is established with a two-level difference modeling paradigm. Specifically, for local motion modeling, temporal difference over consecutive frames is used to supply 2D CNNs with finer motion pattern, while for global motion modeling, temporal difference across segments is incorporated to capture long-range structure for motion feature excitation. TDN provides a simple and principled temporal modeling framework and could be instantiated with the existing CNNs at a small extra computational cost. Our TDN presents a new state of the art on the Something-Something V1 & V2 datasets and is on par with the best performance on the Kinetics-400 dataset. In addition, we conduct in-depth ablation studies and plot the visualization results of our TDN, hopefully providing insightful analysis on temporal difference modeling. We release the code at https://github.com/MCG-NJU/TDN.

Yuan Zhi, Zhan Tong, Limin Wang et al.

Frame sampling is a fundamental problem in video action recognition due to the essential redundancy in time and limited computation resources. The existing sampling strategy often employs a fixed frame selection and lacks the flexibility to deal with complex variations in videos. In this paper, we present a simple, sparse, and explainable frame sampler, termed as Motion-Guided Sampler (MGSampler). Our basic motivation is that motion is an important and universal signal that can drive us to adaptively select frames from videos. Accordingly, we propose two important properties in our MGSampler design: motion sensitive and motion uniform. First, we present two different motion representations to enable us to efficiently distinguish the motion-salient frames from the background. Then, we devise a motion-uniform sampling strategy based on the cumulative motion distribution to ensure the sampled frames evenly cover all the important segments with high motion salience. Our MGSampler yields a new principled and holistic sampling scheme, that could be incorporated into any existing video architecture. Experiments on five benchmarks demonstrate the effectiveness of our MGSampler over the previous fixed sampling strategies, and its generalization power across different backbones, video models, and datasets.