Referring Video Segmentation (RVOS) aims to segment objects in videos given linguistic expressions. The key to solving RVOS is to extract long-range temporal context information from the interactions of expressions and videos to depict the dynamic attributes of each object. Previous works either adopt attention across all the frames or stack dense local attention to achieve a global view of temporal context. However, they fail to strike a good balance between locality and globality, and the computation complexity significantly increases with the increase of video length. In this paper, we propose an effective long-range temporal context attention (LTCA) mechanism to aggregate global context information into object features. Specifically, we aggregate the global context information from two aspects. Firstly, we stack sparse local attentions to balance the locality and globality. We design a dilated window attention across frames to aggregate local context information and perform such attention in a stack of layers to enable a global view. Further, we enable each query to attend to a small group of keys randomly selected from a global pool to enhance the globality. Secondly, we design a global query to interact with all the other queries to directly encode the global context information. Experiments show our method achieves new state-of-the-art on four referring video segmentation benchmarks. Notably, our method shows an improvement of 11.3% and 8.3% on the MeViSvaluandvaldatasets respectively. The code is available at https://github.com/cilinyan/LTCA.

Cilin Yan, Jingyun Wang, Lin Zhang, Ruihui Zhao, Xiaopu Wu, Kai Xiong, Qingsong Liu, Guoliang Kang, Yangyang Kang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Large Vision-Language Models (LVLMs) usually generate texts which satisfy context coherence but don't match the visual input. Such a hallucination issue hinders LVLMs' applicability in the real world. The key to solving hallucination in LVLM is to make the text generation rely more on the visual content. Most previous works choose to enhance/adjust the features/output of a specific modality (i.e., visual or textual) to alleviate hallucinations in LVLM, which do not explicitly or systematically enhance the visual reliance. In this paper, we comprehensively investigate the factors that may degenerate the visual reliance in text generation of LVLM from a Bayesian perspective. We propose to mitigate hallucination in LVLM from three aspects. Firstly, we observe that not all visual tokens are informative in generating meaningful texts. We propose to evaluate and remove redundant visual tokens to avoid their disturbance. Secondly, LVLM may encode inappropriate prior information, making it lean toward generating unexpected words. We propose a simple, yet effective way to rectify the prior from a Bayesian perspective. Thirdly, we observe that starting from certain steps, the posterior of next-token prediction conditioned on visual tokens may collapse to a prior distribution which does not depend on any informative visual tokens at all. Thus, we propose to stop further text generation to avoid hallucination. Extensive experiments on three benchmarks, including POPE, CHAIR, and MME, demonstrate that our method can consistently mitigate the hallucination issue of LVLM and performs favorably against previous state-of-the-arts. Codes are available at https://github.com/NeilHnxTcc/EVRB.

Motion customization aims to adapt the diffusion model (DM) to generate videos with the motion specified by a set of video clips with the same motion concept. To realize this goal, the adaptation of DM should be possible to model the specified motion concept, without compromising the ability to generate diverse appearances. Thus, the key to solving this problem lies in how to separate the motion concept from the appearance in the adaptation process of DM. Typical previous works explore different ways to represent and insert a motion concept into large-scale pre-trained text-to-video diffusion models, e.g., learning a motion LoRA, using latent noise residuals, etc. While those methods can encode the motion concept, they also inevitably encode the appearance in reference videos, resulting in weakened appearance generation capability. In this paper, we follow the typical way to learn a motion LoRA to encode the motion concept, but propose two novel strategies to enhance motion-appearance separation, including temporal attention purification (TAP) and appearance highway (AH). Specifically, we assume that in the temporal attention module, the pretrained Value embeddings are sufficient to serve as basic components needed by producing a new motion. Thus, in TAP, we choose only to reshape the temporal attention with motion LoRAs so that Value embeddings can be reorganized to produce a new motion. Further, in AH, we alter the starting point of each skip connection in U-Net from the output of each temporal attention module to the output of each spatial attention module. Extensive experiments demonstrate that compared to previous works, our method can generate videos with appearance more aligned with the text descriptions and motion more consistent with the reference videos.

Class-Incremental Unsupervised Domain Adaptation (CI-UDA) aims to adapt a model from a labeled source domain to an unlabeled target domain, where the sets of potential target classes appearing at different time steps are disjoint and are subsets of the source classes. The key to solving this problem lies in avoiding catastrophic forgetting of knowledge about previous target classes during continuously mitigating the domain shift. Most previous works cumbersomely combine two technical components. On one hand, they need to store and utilize rehearsal target sample from previous time steps to avoid catastrophic forgetting; on the other hand, they perform alignment only between classes shared across domains at each time step. Consequently, the memory will continuously increase and the asymmetric alignment may inevitably result in knowledge forgetting. In this paper, we propose to mine and preserve domain-invariant and class-agnostic knowledge to facilitate the CI-UDA task. Specifically, via using CLIP, we extract the class-agnostic properties which we name as ''attribute''. In our framework, we learn a ''key-value'' pair to represent an attribute, where the key corresponds to the visual prototype and the value is the textual prompt. We maintain two attribute dictionaries, each corresponding to a different domain. Then we perform attribute alignment across domains to mitigate the domain shift, via encouraging visual attention consistency and prediction consistency. Through attribute modeling and cross-domain alignment, we effectively reduce catastrophic knowledge forgetting while mitigating the domain shift, in a rehearsal-free way. Experiments on three CI-UDA benchmarks demonstrate that our method outperforms previous state-of-the-art methods and effectively alleviates catastrophic forgetting. Code is available at https://github.com/RyunMi/VisTA.

Naive DDIM inversion process usually suffers from a trajectory deviation issue, i.e., the latent trajectory during reconstruction deviates from the one during inversion. To alleviate this issue, previous methods either learn to mitigate the deviation or design cumbersome compensation strategy to reduce the mismatch error, exhibiting substantial time and computation cost. In this work, we present a nearly free-lunch method (named FreeInv) to address the issue more effectively and efficiently. In FreeInv, we randomly transform the latent representation and keep the transformation the same between the corresponding inversion and reconstruction time-step. It is motivated from a statistical perspective that an ensemble of DDIM inversion processes for multiple trajectories yields a smaller trajectory mismatch error on expectation. Moreover, through theoretical analysis and empirical study, we show that FreeInv performs an efficient ensemble of multiple trajectories. FreeInv can be freely integrated into existing inversion-based image and video editing techniques. Especially for inverting video sequences, it brings more significant fidelity and efficiency improvements. Comprehensive quantitative and qualitative evaluation on PIE benchmark and DAVIS dataset shows that FreeInv remarkably outperforms conventional DDIM inversion, and is competitive among previous state-of-the-art inversion methods, with superior computation efficiency.

Industrial Internet of Things (IIoT) faces significant security challenges such as data privacy and vulnerabilities. Unsupervised anomaly detection aims to identify abnormal patterns by monitoring multivariate time series data of IIoT without anomaly annotation. Previous deep-learning-based methods have high-computation cost, which hinders their deployment in edge devices. In this article, we propose a dynamic anomaly detection network (DADN), which introduces a dynamic anomaly detection mechanism to enable efficient inference. Specifically, a bilateral early-exit mechanism is designed so that each sample can dynamically exit at a certain layer during the forward process to support the anomaly judgement, and the layer where sample exits is adaptively determined at the inference stage. Experimental results show that DADN significantly reduces computational costs and enhances F1 scores in industrial anomaly-detection benchmarks, as shown by a 58.22% decrease in GFLOPS on the SWAT dataset, outperforming previous representative method (anomaly transformer).

In real-world scenarios, the number of training samples across classes usually subjects to a long-tailed distribution. The conventionally trained network may achieve unexpected inferior performance on the rare class compared to the frequent class. Most previous works attempt to rectify the network bias from the data-level or from the classifier-level. Differently, in this paper, we identify that the bias towards the frequent class may be encoded into features, i.e., the rare-specific features which play a key role in discriminating the rare class are much weaker than the frequent-specific features. Based on such an observation, we introduce a simple yet effective approach, normalizing the parameters of Batch Normalization (BN) layer to explicitly rectify the feature bias. To achieve this end, we represent theWeight/Bias parameters of a BN layer as a vector, normalize it into a unit one and multiply the unit vector by a scalar learnable parameter. Through decoupling the direction and magnitude of parameters in BN layer to learn, the Weight/Bias exhibits a more balanced distribution and thus the strength of features becomes more even. Extensive experiments on various long-tailed recognition benchmarks (i.e., CIFAR-10/100-LT, ImageNet-LT and iNaturalist 2018) show that our method outperforms previous state-of-the-arts remarkably.

Previous methods utilize the Neural Radiance Field (NeRF) for panoptic lifting, while their training and rendering speed are unsatisfactory. In contrast, 3D Gaussian Splatting (3DGS) has emerged as a prominent technique due to its rapid training and rendering speed. However, unlike NeRF, the conventional 3DGS may not satisfy the basic smoothness assumption as it does not rely on any parameterized structures to render (e.g., MLPs). Consequently, the conventional 3DGS is, in nature, more susceptible to noisy 2D mask supervision. In this paper, we propose a new method called PLGS that enables 3DGS to generate consistent panoptic segmentation masks from noisy 2D segmentation masks while maintaining superior efficiency compared to NeRF-based methods. Specifically, we build a panoptic-aware structured 3D Gaussian model to introduce smoothness and design effective noise reduction strategies. For the semantic field, instead of initialization with structure from motion, we construct reliable semantic anchor points to initialize the 3D Gaussians. We then use these anchor points as smooth regularization during training. Additionally, we present a self-training approach using pseudo labels generated by merging the rendered masks with the noisy masks to enhance the robustness of PLGS. For the instance field, we project the 2D instance masks into 3D space and match them with oriented bounding boxes to generate cross-view consistent instance masks for supervision. Experiments on various benchmarks demonstrate that our method outperforms previous state-of-the-art methods in terms of both segmentation quality and speed.

Consistent editing of real images is a challenging task, as it requires performing non-rigid edits (e.g., changing postures) to the main objects in the input image without changing their identity or attributes. To guarantee consistent attributes, some existing methods fine-tune the entire model or the textual embedding for structural consistency, but they are time-consuming and fail to perform non-rigid edits. Other works are tuning-free, but their performances are weakened by the quality of Denoising Diffusion Implicit Model (DDIM) reconstruction, which often fails in real-world scenarios. In this paper, we present a novel approach called Tuning-free Inversion-enhanced Control (TIC), which directly correlates features from the inversion process with those from the sampling process to mitigate the inconsistency in DDIM reconstruction. Specifically, our method effectively obtains inversion features from the key and value features in the self-attention layers, and enhances the sampling process by these inversion features, thus achieving accurate reconstruction and content-consistent editing. To extend the applicability of our method to general editing scenarios, we also propose a mask-guided attention concatenation strategy that combines contents from both the inversion and the naive DDIM editing processes. Experiments show that the proposed method outperforms previous works in reconstruction and consistent editing, and produces impressive results in various settings.

Recent works utilize CLIP to perform the challenging unsupervised semantic segmentation task where only images without annotations are available. However, we observe that when adopting CLIP to such a pixel-level understanding task, unexpected bias occurs. Previous works don't explicitly model such bias, which largely constrains the segmentation performance. In this paper, we propose to explicitly model and rectify the bias existing in CLIP to facilitate the unsupervised semantic segmentation. Specifically, we design a learnable “Reference” prompt to encode class-preference bias and project the positional embedding of vision transformer to represent space-preference bias. Via a simple element-wise subtraction, we rectify the logits of CLIP classifier. Based on the rectified logits, we generate a segmentation mask via a Gumbel-Softmax operation. Then a contrastive loss between masked visual feature and the text features of different classes is imposed to facilitate the effective bias modeling. To further improve the segmentation, we distill the knowledge from the rectified CLIP to the advanced segmentation architecture via minimizing our designed mask-guided, feature-guided and text-guided loss terms. Extensive experiments on standard benchmarks demonstrate that our method performs favorably against previous state-of-the-arts. The implementation is available at https://github.com/dogehhh/ReCLIP.

Unsupervised domain adaptive person reidentification (UDA re-ID) aims to transfer knowledge learned from from a source domain to a target domain. Prevalent clustering-based self-training methods suffer from label noise, which impedes improvements of model performance. To overcome this issue, mutual networks are introduced to generate reliable soft pseudo-labels for supervising the model training, allowing the suppression of label noise through the complementarity of two networks in the framework. However, existing methods based on this framework face two issues. One issue is that the complementarity between the two networks diminishes as training progresses, reducing the effectiveness of noise suppression. The other issue is that the reliability of pseudo-labels from clustering is not explicitly evaluated, and directly using all pseudo-labels for network optimization may introduce numerous noisy samples. To address these issues, we propose the Data Complementary Training Framework (DCTF) based on mutual networks. This framework comprises two main strategies. Firstly, we introduce the Camera Complementary Training Method (CCTM). CCTM utilizes camera style transfer to generate images of different camera styles for training the two networks, thereby increasing their diversity and enhancing the complementarity of mutual networks. Secondly, we propose the pseudo-labels Reliability Evaluation based on camera Style Transfer (REST). REST assesses the pseudo-label reliability based on the data distribution characteristics, selecting more reliable samples for network optimization. Experimental results on four UDA re-ID tasks demonstrate the effectiveness of our approach, with an mAP gain up to 8.7% compared to previous state-of-the-art methods on the challenging MSMT17 dataset.

Inspired by the recent progress in object detection (i.e., DETR), the set prediction mechanism significantly advances the research of semantic segmentation and achieves state-of-the-art performance on popular segmentation benchmarks. The generic pipeline of such a mechanism often firstly takes learnable query features to predict classes and segment masks separately and then blends these class-aware segment masks into the final segmentation mask. One key factor behind the successful training of this pipeline is to apply thebipartite matchingstrategy between the set of predictions and ground-truth segments. However, we find that thebipartite matching-based assignment often tends to segment one target class with only a few learnable queries, making many other pre-defined queries useless. In this article, we propose a simple way, namedDropQueries(DQ), to facilitate the set prediction based segmentation architectures. At each iteration of training, our DQ randomly and independently drops each learnable query with a certain probability beforebipartite matching. In this way, more queries are encouraged to participate in the segmentation process to discover comprehensive segment representations. We conduct extensive experiments using MaskFormer and Mask2Former as two basic yet powerful segmentation architectures. Without bells and whistles, our DQ strategy can bring consistent improvements over strong baselines on popular semantic segmentation benchmarks, including ADE 20 K, Cityscapes, COCO Stuff 10 K and VSPW.

Domain generalization (DG) aims to train a model with access to a limited number of source domains for generalizing it across various unseen target domains. The key to solving the DG problem is disentangling domain-invariant features (i.e., semantic factors) from domain-specific features (i.e., variation factors) to facilitate generalizable representation learning. Previous studies either implicitly model the semantic and variation factors or ineffectively constrain the disentangling process, thus rendering the disentanglement incomplete and ineffective. In this study, we propose a novel approach, namedDualVAE, to explicitly model and disentangle both the semantic and variation factors. DualVAE is based on the variational autoencoder (VAE) architecture. However, it differs from the conventional VAE in that it consists of two paths, which explicitly model the semantic and variation factors. In addition to the reconstruction loss of VAE and the classification loss, three types of regularizations, namely statistical independence regularization, factorized prior regularization, and prediction consistency regularization, are proposed to further facilitate the disentanglement of factors. Experimental results on representative DG benchmarks show that our method performs favourably against previous state-of-the-art methods. Ablation and visualization results demonstrate that semantic and variation factors can be effectively disentangled.

This paper considers the synthetic-to-real adaptation of point cloud semantic segmentation, which aims to segment the real-world point clouds with only synthetic labels available. Contrary to synthetic data which is integral and clean, point clouds collected by real-world sensors typically contain unexpected and irregular noise because the sensors may be impacted by various environmental conditions. Consequently, the model trained on ideal synthetic data may fail to achieve satisfactory segmentation results on real data. Influenced by such noise, previous adversarial training methods, which are conventional for 2D adaptation tasks, become less effective. In this paper, we aim to mitigate the domain gap caused by target noise via learning to mask the source points during the adaptation procedure. To this end, we design a novel learnable masking module, which takes source features and 3D coordinates as inputs. We incorporate Gumbel-Softmax operation into the masking module so that it can generate binary masks and be trained end-to-end via gradient back-propagation. With the help of adversarial training, the masking module can learn to generate source masks to mimic the pattern of irregular target noise, thereby narrowing the domain gap. We name our method “Adversarial Masking” as adversarial training and learnable masking module depend on each other and cooperate with each other to mitigate the domain gap. Experiments on two synthetic-to-real adaptation benchmarks verify the effectiveness of the proposed method.

Continual learning aims to enable a model to incrementally learn knowledge from sequentially arrived data. Previous works adopt the conventional classification architecture, which consists of a feature extractor and a classifier. The feature extractor is shared across sequentially arrived tasks or classes, but one specific group of weights of the classifier corresponding to one new class should be incrementally expanded. Consequently, the parameters of a continual learner gradually increase. Moreover, as the classifier contains all historical arrived classes, a certain size of the memory is usually required to store rehearsal data to mitigate classifier bias and catastrophic forgetting. In this paper, we propose a non-incremental learner, named AttriCLIP, to incrementally extract knowledge of new classes or tasks. Specifically, AttriCLIP is built upon the pre-trained visual-language model CLIP. Its image encoder and text encoder are fixed to extract features from both images and text. Text consists of a category name and a fixed number of learnable parameters which are selected from our designed attribute word bank and serve as attributes. As we compute the visual and textual similarity for classification, AttriCLIP is a non-incremental learner. The attribute prompts, which encode the common knowledge useful for classification, can effectively mitigate the catastrophic forgetting and avoid constructing a replay memory. We evaluate our AttriCLIP and compare it with CLIP-based and previous state-of-the-art continual learning methods in realistic settings with domain-shift and long-sequence learning. The results show that our method performs favorably against previous state-of-the-arts. The implementation code will be available at https://gitee.com/mindspore/models/tree/master/research/cv/AttriCLIP.

The goal of continual learning is to improve the performance of recognition models in learning sequentially arrived data. Although most existing works are established on the premise of learning from scratch, growing efforts have been devoted to incorporating the benefits of pre-training. However, how to adaptively exploit the pre-trained knowledge for each incremental task while maintaining its generalizability remains an open question. In this work, we present an extensive analysis for continual learning on a pre-trained model (CLPM), and attribute the key challenge to a progressive overfitting problem. Observing that selectively reducing the learning rate can almost resolve this issue in the representation layer, we propose a simple but extremely effective approach named Slow Learner with Classifier Alignment (SLCA), which further improves the classification layer by modeling the class-wise distributions and aligning the classification layers in a post-hoc fashion. Across a variety of scenarios, our proposal provides substantial improvements for CLPM (e.g., up to 49.76%, 50.05%, 44.69% and 40.16% on Split CIFAR-100, Split ImageNet-R, Split CUB-200 and Split Cars-196, respectively), and thus outperforms state-of-the-art approaches by a large margin. Based on such a strong baseline, critical factors and promising directions are analyzed in-depth to facilitate subsequent research. Code has been made available at: https://github.com/GengDavid/SLCA.

Unsupervised domain adaptation (UDA) makes predictions for the target domain data while manual annotations are only available in the source domain. Previous methods minimize the domain discrepancy neglecting the class information, which may lead to misalignment and poor generalization performance. To tackle this issue, this paper proposes contrastive adaptation network (CAN) that optimizes a new metric named Contrastive Domain Discrepancy explicitly modeling the intra-class domain discrepancy and the inter-class domain discrepancy. To optimize CAN, two technical issues need to be addressed: 1) the target labels are not available; and 2) the conventional mini-batch sampling is imbalanced. Thus we design an alternating update strategy to optimize both the target label estimations and the feature representations. Moreover, we develop class-aware sampling to enable more efficient and effective training. Our framework can be generally applied to the single-source and multi-source domain adaptation scenarios. In particular, to deal with multiple source domain data, we propose: 1) multi-source clustering ensemble which exploits the complementary knowledge of distinct source domains to make more accurate and robust target label estimations; and 2) boundary-sensitive alignment to make the decision boundary better fitted to the target. Experiments are conducted on three real-world benchmarks (i.e., Office-31 and VisDA-2017 for the single-source scenario, DomainNet for the multi-source scenario). All the results demonstrate that our CAN performs favorably against the state-of-the-art methods. Ablation studies also verify the effectiveness of each key component of our proposed system.

In this paper, we investigate Universal Domain Adaptation (UniDA) problem, which aims to transfer the knowledge from source to target under unaligned label space. The main challenge of UniDA lies in how to separate common classes (i.e., classes shared across domains), from private classes (i.e., classes only exist in one domain). Previous works treat the private samples in the target as one generic class but ignore their intrinsic structure. Consequently, the resulting representations are not compact enough in the latent space and can be easily confused with common samples. To better exploit the intrinsic structure of the target domain, we propose Domain Consensus Clustering (DCC), which exploits the domain consensus knowledge to discover discriminative clusters on both common samples and private ones. Specifically, we draw the domain consensus knowledge from two aspects to facilitate the clustering and the private class discovery, i.e., the semantic-level consensus, which identifies the cycle-consistent clusters as the common classes, and the sample-level consensus, which utilizes the cross-domain classification agreement to determine the number of clusters and discover the private classes. Based on DCC, we are able to separate the private classes from the common ones, and differentiate the private classes themselves. Finally, we apply a class-aware alignment technique on identified common samples to minimize the distribution shift, and a prototypical regularizer to inspire discriminative target clusters. Experiments on four benchmarks demonstrate DCC significantly outperforms previous state-of-the-arts.

Few-shot segmentation aims to train a segmentation model that can fast adapt to novel classes with few exemplars. The conventional training paradigm is to learn to make predictions on query images conditioned on the features from support images. Previous methods only utilized the semantic-level prototypes of support images as the conditional information. These methods cannot utilize all pixel-wise support information for the query predictions, which is however critical for the segmentation task. In this paper, we focus on utilizing pixel-wise relationships between support and target images to facilitate the few-shot semantic segmentation task. We design a novel Cycle-Consistent Transformer (CyCTR) module to aggregate pixel-wise support features into query ones. CyCTR performs cross-attention between features from different images, i.e. support and query images. We observe that there may exist unexpected irrelevant pixel-level support features. Directly performing cross-attention may aggregate these features from support to query and bias the query features. Thus, we propose using a novel cycle-consistent attention mechanism to filter out possible harmful support features and encourage query features to attend to the most informative pixels from support images. Experiments on all few-shot segmentation benchmarks demonstrate that our proposed CyCTR leads to remarkable improvement compared to previous state-of-the-art methods. Specifically, on Pascal-5^i and COCO-20^i datasets, we achieve 66.6% and 45.6% mIoU for 5-shot segmentation, outperforming previous state-of-the-art by 4.6% and 7.1% respectively.

In this paper, we introduce Random Erasing, a new data augmentation method for training the convolutional neural network (CNN). In training, Random Erasing randomly selects a rectangle region in an image and erases its pixels with random values. In this process, training images with various levels of occlusion are generated, which reduces the risk of over-fitting and makes the model robust to occlusion. Random Erasing is parameter learning free, easy to implement, and can be integrated with most of the CNN-based recognition models. Albeit simple, Random Erasing is complementary to commonly used data augmentation techniques such as random cropping and flipping, and yields consistent improvement over strong baselines in image classification, object detection and person re-identification. Code is available at: https://github.com/zhunzhong07/Random-Erasing.

Domain adaptive semantic segmentation aims to train a model performing satisfactory pixel-level predictions on the target with only out-of-domain (source) annotations. The conventional solution to this task is to minimize the discrepancy between source and target to enable effective knowledge transfer. Previous domain discrepancy minimization methods are mainly based on the adversarial training. They tend to consider the domain discrepancy globally, which ignore the pixel-wise relationships and are less discriminative. In this paper, we propose to build the pixel-level cycle association between source and target pixel pairs and contrastively strengthen their connections to diminish the domain gap and make the features more discriminative. To the best of our knowledge, this is a new perspective for tackling such a challenging task. Experiment results on two representative domain adaptation benchmarks, i.e. GTAV $\rightarrow$ Cityscapes and SYNTHIA $\rightarrow$ Cityscapes, verify the effectiveness of our proposed method and demonstrate that our method performs favorably against previous state-of-the-arts. Our method can be trained end-to-end in one stage and introduce no additional parameters, which is expected to serve as a general framework and help ease future research in domain adaptive semantic segmentation. Code is available at https://github.com/kgl-prml/Pixel-Level-Cycle-Association.

Deeper and wider convolutional neural networks (CNNs) achieve superior performance but bring expensive computation cost. Accelerating such overparameterized neural network has received increased attention. A typical pruning algorithm is a three-stage pipeline, i.e., training, pruning, and retraining. Prevailing approaches fix the pruned filters to zero during retraining and, thus, significantly reduce the optimization space. Besides, they directly prune a large number of filters at first, which would cause unrecoverable information loss. To solve these problems, we propose an asymptotic soft filter pruning (ASFP) method to accelerate the inference procedure of the deep neural networks. First, we update the pruned filters during the retraining stage. As a result, the optimization space of the pruned model would not be reduced but be the same as that of the original model. In this way, the model has enough capacity to learn from the training data. Second, we prune the network asymptotically. We prune few filters at first and asymptotically prune more filters during the training procedure. With asymptotic pruning, the information of the training set would be gradually concentrated in the remaining filters, so the subsequent training and pruning process would be stable. The experiments show the effectiveness of our ASFP on image classification benchmarks. Notably, on ILSVRC-2012, our ASFP reduces more than 40% FLOPs on ResNet-50 with only 0.14% top-5 accuracy degradation, which is higher than the soft filter pruning by 8%.

We focus on the problem of detecting traffic events in a surveillance scenario, including the detection of both vehicle actions and traffic collisions. Existing event detection systems are mostly learning-based and have achieved convincing performance when a large amount of training data is available. However, in real-world scenarios, collecting sufficient labeled training data is expensive and sometimes impossible (e.g. for traffic collision detection). Moreover, the conventional 2D representation of surveillance views is easily affected by occlusions and different camera views in nature. To deal with the aforementioned problems, in this paper, we propose a training-free monocular 3D event detection system for traffic surveillance. Our system firstly projects the vehicles into the 3D Euclidean space and estimates their kinematic states. Then we develop multiple simple yet effective ways to identify the events based on the kinematic patterns, which need no further training. Consequently, our system is robust to the occlusions and the viewpoint changes. Exclusive experiments report the superior result of our method on large-scale real-world surveillance datasets, which validates the effectiveness of our proposed system. The demonstration videos of our system are available online1.

Industrial recommender systems have embraced deep learning algorithms for building intelligent systems to make accurate recommendations. At its core, deep learning offers powerful ability for learning representations from data, especially for user and item representations. Existing deep learning-based models usually represent a user by one representation vector, which is usually insufficient to capture diverse interests for large-scale users in practice. In this paper, we approach the learning of user representations from a different view, by representing a user with multiple representation vectors encoding the different aspects of the user's interests. To this end, we propose the Multi-Interest Network with Dynamic routing (MIND) for learning user representations in recommender systems. Specifically, we design a multi-interest extractor layer based on the recently proposed dynamic routing mechanism, which is applicable for modeling and extracting diverse interests from user's behaviors. Furthermore, a technique named label-aware attention is proposed to help the learning process of user representations. Through extensive experiments on several public benchmarks and one large-scale industrial dataset from Tmall, we demonstrate that MIND can achieve superior performance than state-of-the-art methods in terms of recommendation accuracy. Currently, MIND has been deployed for handling major online traffic at the homepage on Mobile Tmall App.

Unsupervised Domain Adaptation (UDA) makes predictions for the target domain data while manual annotations are only available in the source domain. Previous methods minimize the domain discrepancy neglecting the class information, which may lead to misalignment and poor generalization performance. To address this issue, this paper proposes Contrastive Adaptation Network (CAN) optimizing a new metric which explicitly models the intra-class domain discrepancy and the inter-class domain discrepancy. We design an alternating update strategy for training CAN in an end-to-end manner. Experiments on two real-world benchmarks Office-31 and VisDA-2017 demonstrate that CAN performs favorably against the state-of-the-art methods and produces more discriminative features.

Open set domain adaptation aims to diminish the domain shift across domains, with partially shared classes. There exist unknown target samples out of the knowledge of source domain. Compared to the close set setting, how to separate the unknown (unshared) class from the known (shared) ones plays the key role. Whereas, previous methods did not emphasize the semantic structure of the open set data, which may introduce bias into the domain alignment and confuse the classifier around the decision boundary. In this paper, we exploit the semantic structure of open set data from two aspects: 1) Semantic Categorical Alignment, which aims to achieve good separability of target known classes by categorically aligning the centroid of target with the source. 2) Semantic Contrastive Mapping, which aims to push the unknown class away from the decision boundary. Empirically, we demonstrate that our method performs favourably against the state-of-the-art methods on representative benchmarks, e.g. Digits and Office-31 datasets.

Visual-semantic embeddings are central to many multimedia applications such as cross-modal retrieval between visual data and natural language descriptions. Conventionally, learning a joint embedding space relies on large parallel multimodal corpora. Since massive human annotation is expensive to obtain, there is a strong motivation in developing versatile algorithms to learn from large corpora with fewer annotations. In this paper, we propose a novel framework to leverage automatically extracted regional semantics from un-annotated images as additional weak supervision to learn visual-semantic embeddings. The proposed model employs adversarial attentive alignments to close the inherent heterogeneous gaps between annotated and un-annotated portions of visual and textual domains. To demonstrate its superiority, we conduct extensive experiments on sparsely annotated multimodal corpora. The experimental results show that the proposed model outperforms state-of-the-art visual-semantic embedding models by a significant margin for cross-modal retrieval tasks on the sparse Flickr30k and MS-COCO datasets. It is also worth noting that, despite using only 20% of the annotations, the proposed model can achieve competitive performance (Recall at 10 > 80.0% for 1K and > 70.0% for 5K text-to-image retrieval) compared to the benchmarks trained with the complete annotations.

Person re-identification (re-ID) models trained on one domain often fail to generalize well to another. In our attempt, we present a "learning via translation" framework. In the baseline, we translate the labeled images from source to target domain in an unsupervised manner. We then train re-ID models with the translated images by supervised methods. Yet, being an essential part of this framework, unsupervised image-image translation suffers from the information loss of source-domain labels during translation. Our motivation is two-fold. First, for each image, the discriminative cues contained in its ID label should be maintained after translation. Second, given the fact that two domains have entirely different persons, a translated image should be dissimilar to any of the target IDs. To this end, we propose to preserve two types of unsupervised similarities, 1) self-similarity of an image before and after translation, and 2) domain-dissimilarity of a translated source image and a target image. Both constraints are implemented in the similarity preserving generative adversarial network (SPGAN) which consists of an Siamese network and a CycleGAN. Through domain adaptation experiment, we show that images generated by SPGAN are more suitable for domain adaptation and yield consistent and competitive re-ID accuracy on two large-scale datasets.

This paper proposed a Soft Filter Pruning (SFP) method to accelerate the inference procedure of deep Convolutional Neural Networks (CNNs). Specifically, the proposed SFP enables the pruned filters to be updated when training the model after pruning. SFP has two advantages over previous works: (1) Larger model capacity. Updating previously pruned filters provides our approach with larger optimization space than fixing the filters to zero. Therefore, the network trained by our method has a larger model capacity to learn from the training data. (2) Less dependence on the pretrained model. Large capacity enables SFP to train from scratch and prune the model simultaneously. In contrast, previous filter pruning methods should be conducted on the basis of the pre-trained model to guarantee their performance. Empirically, SFP from scratch outperforms the previous filter pruning methods. Moreover, our approach has been demonstrated effective for many advanced CNN architectures. Notably, on ILSCRC-2012, SFP reduces more than 42% FLOPs on ResNet-101 with even 0.2% top-5 accuracy improvement, which has advanced the state-of-the-art. Code is publicly available on GitHub: https://github.com/he-y/softfilter-pruning

Recent years have witnessed the success of deep neural networks in dealing with a plenty of practical problems. Dropout has played an essential role in many successful deep neural networks, by inducing regularization in the model training. In this paper, we present a new regularized training approach: Shakeout. Instead of randomly discarding units as Dropout does at the training stage, Shakeout randomly chooses to enhance or reverse each unit's contribution to the next layer. This minor modification of Dropout has the statistical trait: the regularizer induced by Shakeout adaptively combines , and regularization terms. Our classification experiments with representative deep architectures on image datasets MNIST, CIFAR-10 and ImageNet show that Shakeout deals with over-fitting effectively and outperforms Dropout. We empirically demonstrate that Shakeout leads to sparser weights under both unsupervised and supervised settings. Shakeout also leads to the grouping effect of the input units in a layer. Considering the weights in reflecting the importance of connections, Shakeout is superior to Dropout, which is valuable for the deep model compression. Moreover, we demonstrate that Shakeout can effectively reduce the instability of the training process of the deep architecture.

Recent years have witnessed the success of deep neural networks in dealing with a plenty of practical problems. The invention of effective training techniques largely contributes to this success. The so-called "Dropout" training scheme is one of the most powerful tool to reduce over-fitting. From the statistic point of view, Dropout works by implicitly imposing an L2 regularizer on the weights. In this paper, we present a new training scheme: Shakeout. Instead of randomly discarding units as Dropout does at the training stage, our method randomly chooses to enhance or inverse the contributions of each unit to the next layer. We show that our scheme leads to a combination of L1 regularization and L2 regularization imposed on the weights, which has been proved effective by the Elastic Net models in practice.We have empirically evaluated the Shakeout scheme and demonstrated that sparse network weights are obtained via Shakeout training. Our classification experiments on real-life image datasets MNIST and CIFAR-10 show that Shakeout deals with over-fitting effectively.

Semi-supervised image classification aims to classify a large quantity of unlabeled images by typically harnessing scarce labeled images. Existing semi-supervised methods often suffer from inadequate classification accuracy when encountering difficult yet critical images, such as outliers, because they treat all unlabeled images equally and conduct classifications in an imperfectly ordered sequence. In this paper, we employ the curriculum learning methodology by investigating the difficulty of classifying every unlabeled image. The reliability and the discriminability of these unlabeled images are particularly investigated for evaluating their difficulty. As a result, an optimized image sequence is generated during the iterative propagations, and the unlabeled images are logically classified from simple to difficult. Furthermore, since images are usually characterized by multiple visual feature descriptors, we associate each kind of features with a teacher, and design a multi-modal curriculum learning (MMCL) strategy to integrate the information from different feature modalities. In each propagation, each teacher analyzes the difficulties of the currently unlabeled images from its own modality viewpoint. A consensus is subsequently reached among all the teachers, determining the currently simplest images (i.e., a curriculum), which are to be reliably classified by the multi-modal learner. This well-organized propagation process leveraging multiple teachers and one learner enables our MMCL to outperform five state-of-the-art methods on eight popular image data sets.