Xiao Liu

Aohan Zeng, Xiao Liu, Zhengxiao Du et al. · tsinghua

We introduce GLM-130B, a bilingual (English and Chinese) pre-trained language model with 130 billion parameters. It is an attempt to open-source a 100B-scale model at least as good as GPT-3 (davinci) and unveil how models of such a scale can be successfully pre-trained. Over the course of this effort, we face numerous unexpected technical and engineering challenges, particularly on loss spikes and divergence. In this paper, we introduce the training process of GLM-130B including its design choices, training strategies for both efficiency and stability, and engineering efforts. The resultant GLM-130B model offers significant outperformance over GPT-3 175B (davinci) on a wide range of popular English benchmarks while the performance advantage is not observed in OPT-175B and BLOOM-176B. It also consistently and significantly outperforms ERNIE TITAN 3.0 260B -- the largest Chinese language model -- across related benchmarks. Finally, we leverage a unique scaling property of GLM-130B to reach INT4 quantization without post training, with almost no performance loss, making it the first among 100B-scale models and more importantly, allowing its effective inference on 4$\times$RTX 3090 (24G) or 8$\times$RTX 2080 Ti (11G) GPUs, the most affordable GPUs required for using 100B-scale models. The GLM-130B model weights are publicly accessible and its code, training logs, related toolkit, and lessons learned are open-sourced at \url{https://github.com/THUDM/GLM-130B/}.

Bohan Li, Ye Yuan, Dingkang Liang et al.

Recently, most handwritten mathematical expression recognition (HMER) methods adopt the encoder-decoder networks, which directly predict the markup sequences from formula images with the attention mechanism. However, such methods may fail to accurately read formulas with complicated structure or generate long markup sequences, as the attention results are often inaccurate due to the large variance of writing styles or spatial layouts. To alleviate this problem, we propose an unconventional network for HMER named Counting-Aware Network (CAN), which jointly optimizes two tasks: HMER and symbol counting. Specifically, we design a weakly-supervised counting module that can predict the number of each symbol class without the symbol-level position annotations, and then plug it into a typical attention-based encoder-decoder model for HMER. Experiments on the benchmark datasets for HMER validate that both joint optimization and counting results are beneficial for correcting the prediction errors of encoder-decoder models, and CAN consistently outperforms the state-of-the-art methods. In particular, compared with an encoder-decoder model for HMER, the extra time cost caused by the proposed counting module is marginal. The source code is available at https://github.com/LBH1024/CAN.

Xiao Liu, Da Yin, Yansong Feng et al. · pku

Spatial commonsense, the knowledge about spatial position and relationship between objects (like the relative size of a lion and a girl, and the position of a boy relative to a bicycle when cycling), is an important part of commonsense knowledge. Although pretrained language models (PLMs) succeed in many NLP tasks, they are shown to be ineffective in spatial commonsense reasoning. Starting from the observation that images are more likely to exhibit spatial commonsense than texts, we explore whether models with visual signals learn more spatial commonsense than text-based PLMs. We propose a spatial commonsense benchmark that focuses on the relative scales of objects, and the positional relationship between people and objects under different actions. We probe PLMs and models with visual signals, including vision-language pretrained models and image synthesis models, on this benchmark, and find that image synthesis models are more capable of learning accurate and consistent spatial knowledge than other models. The spatial knowledge from image synthesis models also helps in natural language understanding tasks that require spatial commonsense.

Xiao Liu, Peng Gao, Tao Yu et al.

Deep learning, especially convolutional neural networks (CNNs) and Transformer architectures, have become the focus of extensive research in medical image segmentation, achieving impressive results. However, CNNs come with inductive biases that limit their effectiveness in more complex, varied segmentation scenarios. Conversely, while Transformer-based methods excel at capturing global and long-range semantic details, they suffer from high computational demands. In this study, we propose CSWin-UNet, a novel U-shaped segmentation method that incorporates the CSWin self-attention mechanism into the UNet to facilitate horizontal and vertical stripes self-attention. This method significantly enhances both computational efficiency and receptive field interactions. Additionally, our innovative decoder utilizes a content-aware reassembly operator that strategically reassembles features, guided by predicted kernels, for precise image resolution restoration. Our extensive empirical evaluations on diverse datasets, including synapse multi-organ CT, cardiac MRI, and skin lesions, demonstrate that CSWin-UNet maintains low model complexity while delivering high segmentation accuracy. Codes are available at https://github.com/eatbeanss/CSWin-UNet.

Xiao Liu, Spyridon Thermos, Pedro Sanchez et al.

Learning disentangled representations requires either supervision or the introduction of specific model designs and learning constraints as biases. InfoGAN is a popular disentanglement framework that learns unsupervised disentangled representations by maximising the mutual information between latent representations and their corresponding generated images. Maximisation of mutual information is achieved by introducing an auxiliary network and training with a latent regression loss. In this short exploratory paper, we study the use of the Hilbert-Schmidt Independence Criterion (HSIC) to approximate mutual information between latent representation and image, termed HSIC-InfoGAN. Directly optimising the HSIC loss avoids the need for an additional auxiliary network. We qualitatively compare the level of disentanglement in each model, suggest a strategy to tune the hyperparameters of HSIC-InfoGAN, and discuss the potential of HSIC-InfoGAN for medical applications.

Xiao Liu, Guangyi Chen, Yansong Tang et al.

Composing simple elements into complex concepts is crucial yet challenging, especially for 3D action generation. Existing methods largely rely on extensive neural language annotations to discern composable latent semantics, a process that is often costly and labor-intensive. In this study, we introduce a novel framework to generate compositional actions without reliance on language auxiliaries. Our approach consists of three main components: Action Coupling, Conditional Action Generation, and Decoupling Refinement. Action Coupling utilizes an energy model to extract the attention masks of each sub-action, subsequently integrating two actions using these attentions to generate pseudo-training examples. Then, we employ a conditional generative model, CVAE, to learn a latent space, facilitating the diverse generation. Finally, we propose Decoupling Refinement, which leverages a self-supervised pre-trained model MAE to ensure semantic consistency between the sub-actions and compositional actions. This refinement process involves rendering generated 3D actions into 2D space, decoupling these images into two sub-segments, using the MAE model to restore the complete image from sub-segments, and constraining the recovered images to match images rendered from raw sub-actions. Due to the lack of existing datasets containing both sub-actions and compositional actions, we created two new datasets, named HumanAct-C and UESTC-C, and present a corresponding evaluation metric. Both qualitative and quantitative assessments are conducted to show our efficacy.

Rui Lu, Zhenyu Hou, Zihan Wang et al.

Augmenting large language models (LLMs) with browsing tools substantially improves their potential as deep search agents to solve complex, real-world tasks. Yet, open LLMs still perform poorly in such settings due to limited long-horizon reasoning capacity with browsing tools and the lack of sufficiently difficult supervised data. To address these challenges, we present DeepDive to advance deep search agents. First, we propose a strategy to automatically synthesize complex, difficult, and hard-to-find questions from open knowledge graphs. Second, we apply end-to-end multi-turn reinforcement learning (RL) to enhance LLMs' long-horizon reasoning with deep search. To encourage diversity and reduce redundancy, we design a redundancy penalty that discourages repeated similar queries. Experiments show that DeepDive-32B achieves a new open-source competitive result on BrowseComp, outperforming WebSailor, DeepSeek-R1-Browse, and Search-o1. We demonstrate that multi-turn RL training improves deep search ability and significantly contributes to the performance improvements across multiple benchmarks. We observe that DeepDive enables test-time scaling of tool calls and parallel sampling. All datasets, models, and code are publicly available at https://github.com/THUDM/DeepDive.

Naitik Khandelwal, Xiao Liu, Mengmi Zhang

Scene graph generation (SGG) analyzes images to extract meaningful information about objects and their relationships. In the dynamic visual world, it is crucial for AI systems to continuously detect new objects and establish their relationships with existing ones. Recently, numerous studies have focused on continual learning within the domains of object detection and image recognition. However, a limited amount of research focuses on a more challenging continual learning problem in SGG. This increased difficulty arises from the intricate interactions and dynamic relationships among objects, and their associated contexts. Thus, in continual learning, SGG models are often required to expand, modify, retain, and reason scene graphs within the process of adaptive visual scene understanding. To systematically explore Continual Scene Graph Generation (CSEGG), we present a comprehensive benchmark comprising three learning regimes: relationship incremental, scene incremental, and relationship generalization. Moreover, we introduce a ``Replays via Analysis by Synthesis" method named RAS. This approach leverages the scene graphs, decomposes and re-composes them to represent different scenes, and replays the synthesized scenes based on these compositional scene graphs. The replayed synthesized scenes act as a means to practice and refine proficiency in SGG in known and unknown environments. Our experimental results not only highlight the challenges of directly combining existing continual learning methods with SGG backbones but also demonstrate the effectiveness of our proposed approach, enhancing CSEGG efficiency while simultaneously preserving privacy and memory usage. All data and source code are publicly available online.

Xiao Liu, Jiawei Zhang

The rapid advancement of AI-generated video models has created a pressing need for robust and interpretable evaluation frameworks. Existing metrics are limited to producing numerical scores without explanatory comments, resulting in low interpretability and human evaluation alignment. To address those challenges, we introduce AIGVE-MACS, a unified model for AI-Generated Video Evaluation(AIGVE), which can provide not only numerical scores but also multi-aspect language comment feedback in evaluating these generated videos. Central to our approach is AIGVE-BENCH 2, a large-scale benchmark comprising 2,500 AI-generated videos and 22,500 human-annotated detailed comments and numerical scores across nine critical evaluation aspects. Leveraging AIGVE-BENCH 2, AIGVE-MACS incorporates recent Vision-Language Models with a novel token-wise weighted loss and a dynamic frame sampling strategy to better align with human evaluators. Comprehensive experiments across supervised and zero-shot benchmarks demonstrate that AIGVE-MACS achieves state-of-the-art performance in both scoring correlation and comment quality, significantly outperforming prior baselines including GPT-4o and VideoScore. In addition, we further showcase a multi-agent refinement framework where feedback from AIGVE-MACS drives iterative improvements in video generation, leading to 53.5% quality enhancement. This work establishes a new paradigm for comprehensive, human-aligned evaluation of AI-generated videos. We release the AIGVE-BENCH 2 and AIGVE-MACS at https://huggingface.co/xiaoliux/AIGVE-MACS.

Zhichao Li, Yi Yang, Xiao Liu et al.

We propose a dynamic computational time model to accelerate the average processing time for recurrent visual attention (RAM). Rather than attention with a fixed number of steps for each input image, the model learns to decide when to stop on the fly. To achieve this, we add an additional continue/stop action per time step to RAM and use reinforcement learning to learn both the optimal attention policy and stopping policy. The modification is simple but could dramatically save the average computational time while keeping the same recognition performance as RAM. Experimental results on CUB-200-2011 and Stanford Cars dataset demonstrate the dynamic computational model can work effectively for fine-grained image recognition.The source code of this paper can be obtained from https://github.com/baidu-research/DT-RAM

Shulong Zhang, Mingyuan Yao, Jiayin Zhao et al.

In recirculating aquaculture systems, accurate and effective assessment of fish feeding intensity is crucial for reducing feed costs and calculating optimal feeding times. However, current studies have limitations in modality selection, feature extraction and fusion, and co-inference for decision making, which restrict further improvement in the accuracy, applicability and reliability of multimodal fusion models. To address this problem, this study proposes a Multi-stage Augmented Multimodal Interaction Network (MAINet) for quantifying fish feeding intensity. Firstly, a general feature extraction framework is proposed to efficiently extract feature information from input image, audio and water wave datas. Second, an Auxiliary-modality Reinforcement Primary-modality Mechanism (ARPM) is designed for inter-modal interaction and generate enhanced features, which consists of a Channel Attention Fusion Network (CAFN) and a Dual-mode Attention Fusion Network (DAFN). Finally, an Evidence Reasoning (ER) rule is introduced to fuse the output results of each modality and make decisions, thereby completing the quantification of fish feeding intensity. The experimental results show that the constructed MAINet reaches 96.76%, 96.78%, 96.79% and 96.79% in accuracy, precision, recall and F1-Score respectively, and its performance is significantly higher than the comparison models. Compared with models that adopt single-modality, dual-modality fusion and different decision-making fusion methods, it also has obvious advantages. Meanwhile, the ablation experiments further verified the key role of the proposed improvement strategy in improving the robustness and feature utilization efficiency of model, which can effectively improve the accuracy of the quantitative results of fish feeding intensity.

Shulong Zhang, Jiayin Zhao, Mingyuan Yao et al.

As a key part of aquaculture management, fish feeding behavior recognition and intensity quantification has been a hot area of great concern to researchers, and it plays a crucial role in monitoring fish health, guiding baiting work and improving aquaculture efficiency. In order to better carry out the related work in the future, this paper firstly analyzes and compares the existing reviews. Then reviews the research advances of fish feeding behavior recognition and intensity quantification methods based on computer vision, acoustics and sensors in a single modality. Meanwhile, the application of the current emerging multimodal fusion in fish feeding behavior recognition and intensity quantification methods is expounded. Finally, the advantages and disadvantages of various techniques are compared and analyzed, and the future research directions are envisioned.

Yuezhou Sun, Wenlong Zhao, Lijun Zhang et al.

This paper investigates deep neural network (DNN) compression from the perspective of compactly representing and storing trained parameters. We explore the previously overlooked opportunity of cross-layer architecture-agnostic representation sharing for DNN parameters. To do this, we decouple feedforward parameters from DNN architectures and leverage additive quantization, an extreme lossy compression method invented for image descriptors, to compactly represent the parameters. The representations are then finetuned on task objectives to improve task accuracy. We conduct extensive experiments on MobileNet-v2, VGG-11, ResNet-50, Feature Pyramid Networks, and pruned DNNs trained for classification, detection, and segmentation tasks. The conceptually simple scheme consistently outperforms iterative unstructured pruning. Applied to ResNet-50 with 76.1% top-1 accuracy on the ILSVRC12 classification challenge, it achieves a $7.2\times$ compression ratio with no accuracy loss and a $15.3\times$ compression ratio at 74.79% accuracy. Further analyses suggest that representation sharing can frequently happen across network layers and that learning shared representations for an entire DNN can achieve better accuracy at the same compression ratio than compressing the model as multiple separate parts. We release PyTorch code to facilitate DNN deployment on resource-constrained devices and spur future research on efficient representations and storage of DNN parameters.

Xiaopeng Bi, Ran Yan, Zheng Chai et al.

This report describes Megvii-3D team's approach towards SimLocMatch Challenge @ CVPR 2021 Image Matching Workshop.

Baoming Yan, Lin Wang, Ke Gao et al.

Video affective understanding, which aims to predict the evoked expressions by the video content, is desired for video creation and recommendation. In the recent EEV challenge, a dense affective understanding task is proposed and requires frame-level affective prediction. In this paper, we propose a multi-granularity network with modal attention (MGN-MA), which employs multi-granularity features for better description of the target frame. Specifically, the multi-granularity features could be divided into frame-level, clips-level and video-level features, which corresponds to visual-salient content, semantic-context and video theme information. Then the modal attention fusion module is designed to fuse the multi-granularity features and emphasize more affection-relevant modals. Finally, the fused feature is fed into a Mixtures Of Experts (MOE) classifier to predict the expressions. Further employing model-ensemble post-processing, the proposed method achieves the correlation score of 0.02292 in the EEV challenge.

Xiao Liu, Xiaofei Si, Jiangtao Xie

This is a short technical report introducing the solution of Team Rat for Short-video Parsing Face Parsing Track of The 3rd Person in Context (PIC) Workshop and Challenge at CVPR 2021. In this report, we propose an Edge-Aware Network (EANet) that uses edge information to refine the segmentation edge. To further obtain the finer edge results, we introduce edge attention loss that only compute cross entropy on the edges, it can effectively reduce the classification error around edge and get more smooth boundary. Benefiting from the edge information and edge attention loss, the proposed EANet achieves 86.16\% accuracy in the Short-video Face Parsing track of the 3rd Person in Context (PIC) Workshop and Challenge, ranked the third place.

Haotian Zhang, Yicheng Luo, Fangbo Qin et al.

We present the novel Efficient Line Segment Detector and Descriptor (ELSD) to simultaneously detect line segments and extract their descriptors in an image. Unlike the traditional pipelines that conduct detection and description separately, ELSD utilizes a shared feature extractor for both detection and description, to provide the essential line features to the higher-level tasks like SLAM and image matching in real time. First, we design the one-stage compact model, and propose to use the mid-point, angle and length as the minimal representation of line segment, which also guarantees the center-symmetry. The non-centerness suppression is proposed to filter out the fragmented line segments caused by lines' intersections. The fine offset prediction is designed to refine the mid-point localization. Second, the line descriptor branch is integrated with the detector branch, and the two branches are jointly trained in an end-to-end manner. In the experiments, the proposed ELSD achieves the state-of-the-art performance on the Wireframe dataset and YorkUrban dataset, in both accuracy and efficiency. The line description ability of ELSD also outperforms the previous works on the line matching task.

Xiao Liu, Jiajie Zhang, Siting Li et al.

What mechanisms causes GAN's entanglement? Although developing disentangled GAN has attracted sufficient attention, it is unclear how entanglement is originated by GAN transformation. We in this research propose a difference-in-difference (DID) counterfactual framework to design experiments for analyzing the entanglement mechanism in on of the Progressive-growing GAN (PG-GAN). Our experiment clarify the mechanisms how pixel normalization causes PG-GAN entanglement during a input-unit-ablation transformation. We discover that pixel normalization causes object entanglement by in-painting the area occupied by ablated objects. We also discover the unit-object relation determines whether and how pixel normalization causes objects entanglement. Our DID framework theoretically guarantees that the mechanisms that we discover is solid, explainable and comprehensively.

Pengcheng Wang, Zihao Wang, Zhilong Ji et al.

This paper introduces our approach to the EmotioNet Challenge 2020. We pose the AU recognition problem as a multi-task learning problem, where the non-rigid facial muscle motion (mainly the first 17 AUs) and the rigid head motion (the last 6 AUs) are modeled separately. The co-occurrence of the expression features and the head pose features are explored. We observe that different AUs converge at various speed. By choosing the optimal checkpoint for each AU, the recognition results are improved. We are able to obtain a final score of 0.746 in validation set and 0.7306 in the test set of the challenge.

Yongcheng Jing, Xiao Liu, Yukang Ding et al.

Prior normalization methods rely on affine transformations to produce arbitrary image style transfers, of which the parameters are computed in a pre-defined way. Such manually-defined nature eventually results in the high-cost and shared encoders for both style and content encoding, making style transfer systems cumbersome to be deployed in resource-constrained environments like on the mobile-terminal side. In this paper, we propose a new and generalized normalization module, termed as Dynamic Instance Normalization (DIN), that allows for flexible and more efficient arbitrary style transfers. Comprising an instance normalization and a dynamic convolution, DIN encodes a style image into learnable convolution parameters, upon which the content image is stylized. Unlike conventional methods that use shared complex encoders to encode content and style, the proposed DIN introduces a sophisticated style encoder, yet comes with a compact and lightweight content encoder for fast inference. Experimental results demonstrate that the proposed approach yields very encouraging results on challenging style patterns and, to our best knowledge, for the first time enables an arbitrary style transfer using MobileNet-based lightweight architecture, leading to a reduction factor of more than twenty in computational cost as compared to existing approaches. Furthermore, the proposed DIN provides flexible support for state-of-the-art convolutional operations, and thus triggers novel functionalities, such as uniform-stroke placement for non-natural images and automatic spatial-stroke control.

Xin Li, Tianwei Lin, Xiao Liu et al.

Existing action localization approaches adopt shallow temporal convolutional networks (\ie, TCN) on 1D feature map extracted from video frames. In this paper, we empirically find that stacking more conventional temporal convolution layers actually deteriorates action classification performance, possibly ascribing to that all channels of 1D feature map, which generally are highly abstract and can be regarded as latent concepts, are excessively recombined in temporal convolution. To address this issue, we introduce a novel concept-wise temporal convolution (CTC) layer as an alternative to conventional temporal convolution layer for training deeper action localization networks. Instead of recombining latent concepts, CTC layer deploys a number of temporal filters to each concept separately with shared filter parameters across concepts. Thus can capture common temporal patterns of different concepts and significantly enrich representation ability. Via stacking CTC layers, we proposed a deep concept-wise temporal convolutional network (C-TCN), which boosts the state-of-the-art action localization performance on THUMOS'14 from 42.8 to 52.1 in terms of mAP(\%), achieving a relative improvement of 21.7\%. Favorable result is also obtained on ActivityNet.

Dongliang He, Xiang Zhao, Jizhou Huang et al.

The task of video grounding, which temporally localizes a natural language description in a video, plays an important role in understanding videos. Existing studies have adopted strategies of sliding window over the entire video or exhaustively ranking all possible clip-sentence pairs in a pre-segmented video, which inevitably suffer from exhaustively enumerated candidates. To alleviate this problem, we formulate this task as a problem of sequential decision making by learning an agent which regulates the temporal grounding boundaries progressively based on its policy. Specifically, we propose a reinforcement learning based framework improved by multi-task learning and it shows steady performance gains by considering additional supervised boundary information during training. Our proposed framework achieves state-of-the-art performance on ActivityNet'18 DenseCaption dataset and Charades-STA dataset while observing only 10 or less clips per video.

Dongliang He, Zhichao Zhou, Chuang Gan et al.

Despite the success of deep learning for static image understanding, it remains unclear what are the most effective network architectures for the spatial-temporal modeling in videos. In this paper, in contrast to the existing CNN+RNN or pure 3D convolution based approaches, we explore a novel spatial temporal network (StNet) architecture for both local and global spatial-temporal modeling in videos. Particularly, StNet stacks N successive video frames into a \emph{super-image} which has 3N channels and applies 2D convolution on super-images to capture local spatial-temporal relationship. To model global spatial-temporal relationship, we apply temporal convolution on the local spatial-temporal feature maps. Specifically, a novel temporal Xception block is proposed in StNet. It employs a separate channel-wise and temporal-wise convolution over the feature sequence of video. Extensive experiments on the Kinetics dataset demonstrate that our framework outperforms several state-of-the-art approaches in action recognition and can strike a satisfying trade-off between recognition accuracy and model complexity. We further demonstrate the generalization performance of the leaned video representations on the UCF101 dataset.

Chen Zhu, Xiao Tan, Feng Zhou et al.

For fine-grained categorization tasks, videos could serve as a better source than static images as videos have a higher chance of containing discriminative patterns. Nevertheless, a video sequence could also contain a lot of redundant and irrelevant frames. How to locate critical information of interest is a challenging task. In this paper, we propose a new network structure, known as Redundancy Reduction Attention (RRA), which learns to focus on multiple discriminative patterns by sup- pressing redundant feature channels. Specifically, it firstly summarizes the video by weight-summing all feature vectors in the feature maps of selected frames with a spatio-temporal soft attention, and then predicts which channels to suppress or to enhance according to this summary with a learned non-linear transform. Suppression is achieved by modulating the feature maps and threshing out weak activations. The updated feature maps are then used in the next iteration. Finally, the video is classified based on multiple summaries. The proposed method achieves out- standing performances in multiple video classification datasets. Further- more, we have collected two large-scale video datasets, YouTube-Birds and YouTube-Cars, for future researches on fine-grained video categorization. The datasets are available at http://www.cs.umd.edu/~chenzhu/fgvc.

Xiao Liu, Shengchuan Zhang, Hong Liu et al.

In this paper, we aim at solving the multi-domain image-to-image translation problem with a unified model in an unsupervised manner. The most successful work in this area refers to StarGAN, which works well in tasks like face attribute modulation. However, StarGAN is unable to match multiple translation mappings when encountering general translations with very diverse domain shifts. On the other hand, StarGAN adopts an Encoder-Decoder-Discriminator (EDD) architecture, where the model is time-consuming and unstable to train. To this end, we propose a Compact, effective, robust, and fast GAN model, termed CerfGAN, to solve the above problem. In principle, CerfGAN contains a novel component, i.e., a multi-class discriminator (MCD), which gives the model an extremely powerful ability to match multiple translation mappings. To stabilize the training process, MCD also plays a role of the encoder in CerfGAN, which saves a lot of computation and memory costs. We perform extensive experiments to verify the effectiveness of the proposed method. Quantitatively, CerfGAN is demonstrated to handle a serial of image-to-image translation tasks including style transfer, season transfer, face hallucination, etc, where the input images are sampled from diverse domains. The comparisons to several recently proposed approaches demonstrate the superiority and novelty of the proposed method.

Xiang Long, Chuang Gan, Gerard de Melo et al.

Recently, substantial research effort has focused on how to apply CNNs or RNNs to better extract temporal patterns from videos, so as to improve the accuracy of video classification. In this paper, however, we show that temporal information, especially longer-term patterns, may not be necessary to achieve competitive results on common video classification datasets. We investigate the potential of a purely attention based local feature integration. Accounting for the characteristics of such features in video classification, we propose a local feature integration framework based on attention clusters, and introduce a shifting operation to capture more diverse signals. We carefully analyze and compare the effect of different attention mechanisms, cluster sizes, and the use of the shifting operation, and also investigate the combination of attention clusters for multimodal integration. We demonstrate the effectiveness of our framework on three real-world video classification datasets. Our model achieves competitive results across all of these. In particular, on the large-scale Kinetics dataset, our framework obtains an excellent single model accuracy of 79.4% in terms of the top-1 and 94.0% in terms of the top-5 accuracy on the validation set. The attention clusters are the backbone of our winner solution at ActivityNet Kinetics Challenge 2017. Code and models will be released soon.