Developing robust and discriminative appearance models has been a long-standing research challenge in visual object tracking. In the prevalent Siamese-based paradigm, the features extracted by the Siamese-like networks are often insufficient to model the tracked targets and distractor objects, thereby hindering them from being robust and discriminative simultaneously. While most Siamese trackers focus on designing robust correlation operations, we propose a novel single-branch tracking framework inspired by the transformer. Unlike the Siamese-like feature extraction, our tracker deeply embeds cross-image feature correlation in multiple layers of the feature network. By extensively matching the features of the two images through multiple layers, it can suppress non-target features, resulting in target-aware feature extraction. The output features can be directly used to predict target locations without additional correlation steps. Thus, we reformulate the two-branch Siamese tracking as a conceptually simple, fully transformer-based Single-Branch Tracking pipeline, dubbed SBT. After conducting an in-depth analysis of the SBT baseline, we summarize many effective design principles and propose an improved tracker dubbed SuperSBT. SuperSBT adopts a hierarchical architecture with a local modeling layer to enhance shallow-level features. A unified relation modeling is proposed to remove complex handcrafted layer pattern designs. SuperSBT is further improved by masked image modeling pre-training, integrating temporal modeling, and equipping with dedicated prediction heads. Thus, SuperSBT outperforms the SBT baseline by 4.7%,3.0%, and 4.5% AUC scores in LaSOT, TrackingNet, and GOT-10K. Notably, SuperSBT greatly raises the speed of SBT from 37 FPS to 81 FPS. Extensive experiments show that our method achieves superior results on eight VOT benchmarks.

Vision transformers (ViT) have shown promise in various vision tasks while the U-Net based on a convolutional neural network (CNN) remains dominant in diffusion models. We design a simple and general ViT-based architecture (named U-ViT) for image generation with diffusion models. U-ViT is characterized by treating all inputs including the time, condition and noisy image patches as tokens and employing long skip connections between shallow and deep layers. We evaluate U-ViT in unconditional and classconditional image generation, as well as text-to-image generation tasks, where U-ViT is comparable if not superior to a CNN-based U-Net of a similar size. In particular, latent diffusion models with U-ViT achieve record-breaking FID scores of 2.29 in class-conditional image generation on ImageNet 256×256, and 5.48 in text-to-image generation on MS-COCO, among methods without accessing large external datasets during the training of generative models. Our results suggest that, for diffusion-based image modeling, the long skip connection is crucial while the down-sampling and upsampling operators in CNN-based U-Net are not always necessary. We believe that U-ViT can provide insights for future research on backbones in diffusion models and benefit generative modeling on large scale cross-modality datasets.

We launch EVA, a vision-centric foundation model to Explore the limits of Visual representation at scAle using only publicly accessible data. EVA is a vanilla ViT pre-trained to reconstruct the masked out image-text aligned vision features conditioned on visible image patches. Via this pretext task, we can efficiently scale up EVA to one billion parameters, and sets new records on a broad range of representative vision downstream tasks, such as image recognition, video action recognition, object detection, instance segmentation and semantic segmentation without heavy supervised training. Moreover, we observe quantitative changes in scaling EVA result in qualitative changes in transfer learning performance that are not present in other models. For instance, EVA takes a great leap in the challenging large vocabulary instance segmentation task: our model achieves almost the same state-of-the-art performance on LVIS dataset with over a thousand categories and COCO dataset with only eighty categories. Beyond a pure vision encoder, EVA can also serve as a vision-centric, multi-modal pivot to connect images and text. We find initializing the vision tower of a giant CLIP from EVA can greatly stabilize the training and outperform the training from scratch counterpart with much fewer samples and less compute, providing a new direction for scaling up and accelerating the costly training of multi-modal foundation models.

In-context learning, as a new paradigm in NLP, allows the model to rapidly adapt to various tasks with only a handful of prompts and examples. But in computer vision, the difficulties for in-context learning lie in that tasks vary significantly in the output representations, thus it is unclear how to define the general-purpose task prompts that the vision model can understand and transfer to out-of-domain tasks. In this work, we present Painter, a generalist model which addresses these obstacles with an “image”-centric solution, that is, to redefine the output of core vision tasks as images, and specify task prompts as also images. With this idea, our training process is extremely simple, which performs standard masked image modeling on the stitch of input and output image pairs. This makes the model capable of performing tasks conditioned on visible image patches. Thus, during inference, we can adopt a pair of input and output images from the same task as the input condition, to indicate which task to perform. Without bells and whistles, our generalist Painter can achieve competitive performance compared to well-established task-specific models, on seven representative vision tasks ranging from high-level visual understanding to low-level image processing. In addition, Painter significantly outperforms recent generalist models on several challenging tasks.

This paper presents a method that effectively combines two prevalent visual recognition methods, i.e., image classification and contrastive language-image pre-training, dubbed iCLIP. Instead of naïve multi-task learning that use two separate heads for each task, we fuse the two tasks in a deep fashion that adapts the image classification to share the same formula and the same model weights with the language-image pre-training. To further bridge these two tasks, we propose to enhance the category names in image classification tasks using external knowledge, such as their descriptions in dictionaries. Extensive experiments show that the proposed method combines the advantages of two tasks well: the strong discrimination ability in image classification tasks due to the clean category labels, and the good zero-shot ability in CLIP tasks ascribed to the richer semantics in the text descriptions. In particular, it reaches 82.9% top-1 accuracy on IN-1K, and mean-while surpasses CLIP by 1.8%, with similar model size, on zero-shot recognition of Kornblith 12-dataset benchmark. The code and models are publicly available at https://github.com/weiyx16/iCLIP.

Masked image modeling (MIM) as pre-training is shown to be effective for numerous vision downstream tasks, but how and where MIM works remain unclear. In this paper, we compare MIM with the long-dominant supervised pretrained models from two perspectives, the visualizations and the experiments, to uncover their key representational differences. From the visualizations, we find that MIM brings locality inductive bias to all layers of the trained models, but supervised models tend to focus locally at lower layers but more globally at higher layers. That may be the reason why MIM helps Vision Transformers that have a very large receptive field to optimize. Using MIM, the model can maintain a large diversity on attention heads in all layers. But for supervised models, the diversity on attention heads almost disappears from the last three layers and less diversity harms the fine-tuning performance. From the experiments, we find that MIM models can perform significantly better on geometric and motion tasks with weak semantics or fine-grained classification tasks, than their supervised counterparts. Without bells and whistles, a standard MIM pre-trained SwinV2-L could achieve state-of-the-art performance on pose estimation (78.9 AP on COCO test-dev and 78.0 AP on CrowdPose), depth estimation (0.287 RMSE on NYUv2 and 1.966 RMSE on KITTI), and video object tracking (70.7 SUC on LaSOT). For the semantic understanding datasets where the categories are sufficiently covered by the supervised pre-training, MIM models can still achieve highly competitive transfer performance. With a deeper understanding of MIM, we hope that our work can inspire new and solid research in this direction. Code will be available at https://github.com/zdaxie/MIM-DarkSecrets.

Scaling properties have been one of the central issues in self-supervised pre-training, especially the data scalability, which has successfully motivated the large-scale self-supervised pre-trained language models and endowed them with significant modeling capabilities. However, scaling properties seem to be unintentionally neglected in the recent trending studies on masked image modeling (MIM), and some arguments even suggest that MIM cannot benefit from large-scale data. In this work, we try to break down these preconceptions and systematically study the scaling behaviors of MIM through extensive experiments, with data ranging from 10% of ImageNet-1K to full ImageNet-22K, model parameters ranging from 49-million to one-billion, and training length ranging from 125K to 500K iterations. And our main findings can be summarized in two folds: 1) masked image modeling remains demanding large-scale data in order to scale up computes and model parameters; 2) masked image modeling cannot benefit from more data under a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-trained language models or supervised pre-trained vision models. In addition, we reveal several intriguing properties in MIM, such as high sample efficiency in large MIM models and strong correlation between pre-training validation loss and transfer performance. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on largescale vision models. Code and models will be available at https://github.com/microsoft/SimMIM.

Recent years have witnessed a remarkable success of large deep learning models. However, training these models is challenging due to high computational costs, painfully slow convergence, and overfitting issues. In this paper, we present Deep Incubation, a novel approach that enables the efficient and effective training of large models by dividing them into smaller sub-modules which can be trained separately and assembled seamlessly. A key challenge for implementing this idea is to ensure the compatibility of the independently trained sub-modules. To address this issue, we first introduce a global, shared meta model, which is leveraged to implicitly link all the modules together, and can be designed as an extremely small network with negligible computational overhead. Then we propose a module incubation algorithm, which trains each sub-module to replace the corresponding component of the meta model and accomplish a given learning task. Despite the simplicity, our approach effectively encourages each sub-module to be aware of its role in the target large model, such that the finally-learned sub-modules can collaborate with each other smoothly after being assembled. Empirically, our method can outperform end-to-end (E2E) training in well-established training setting and shows transferable performance gain for downstream tasks (e.g., object detection and image segmentation on COCO and ADE20K). Our code is available at https://github.com/LeapLabTHU/Deep-Incubation.

We present SegGPT, a generalist model for segmenting everything in context. We unify various segmentation tasks into a generalist in-context learning framework that accommodates different kinds of segmentation data by transforming them into the same format of images. The training of SegGPT is formulated as an in-context coloring problem with random color mapping for each data sample. The objective is to accomplish diverse tasks according to the context, rather than relying on specific colors. After training, SegGPT can perform arbitrary segmentation tasks in images or videos via in-context inference, such as object instance, stuff, part, contour, and text. SegGPT is evaluated on a broad range of tasks, including few-shot semantic segmentation, video object segmentation, semantic segmentation, and panoptic segmentation. Our results show strong capabilities in segmenting in-domain and out-of-domain targets, either qualitatively or quantitatively.

CLIP models have demonstrated impressively high zero-shot recognition accuracy, however, their fine-tuning performance on downstream vision tasks is sub-optimal. Contrarily, masked image modeling (MIM) performs exceptionally for fine-tuning on downstream tasks, despite the absence of semantic labels during training. We note that the two tasks have different ingredients: image-level targets versus token-level targets, a cross-entropy loss versus a regression loss, and full-image inputs versus partial-image inputs. To mitigate the differences, we introduce a classical feature map distillation framework, which can simultaneously inherit the semantic capability of CLIP models while constructing a task incorporated key ingredients of MIM. Experiments suggest that the feature map distillation approach significantly boosts the fine-tuning performance of CLIP models on several typical down-stream vision tasks. We also observe that the approach yields new CLIP representations which share some diagnostic properties with those of MIM. Furthermore, the feature map distillation approach generalizes to other pre-training models, such as DINO, DeiT and SwinV2-G, reaching a new record of 64.2 mAP on COCO object detection with +1.1 improvement. The code and models are publicly available at https://github.com/SwinTransformer/Feature-Distillation.

This paper proposes a unified diffusion framework (dubbed UniDiffuser) to fit all distributions relevant to a set of multi-modal data in one model. Our key insight is – learning diffusion models for marginal, conditional, and joint distributions can be unified as predicting the noise in the perturbed data, where the perturbation levels (i.e. timesteps) can be different for different modalities. Inspired by the unified view, UniDiffuser learns all distributions simultaneously with a minimal modification to the original diffusion model – perturbs data in all modalities instead of a single modality, inputs individual timesteps in different modalities, and predicts the noise of all modalities instead of a single modality. UniDiffuser is parameterized by a transformer for diffusion models to handle input types of different modalities. Implemented on large-scale paired image-text data, UniDiffuser is able to perform image, text, text-to-image, image-to-text, and image-text pair generation by setting proper timesteps without additional overhead. In particular, UniDiffuser is able to produce perceptually realistic samples in all tasks and its quantitative results (e.g., the FID and CLIP score) are not only superior to existing general-purpose models but also comparable to the bespoken models (e.g., Stable Diffusion and DALL-E 2) in representative tasks (e.g., text-to-image generation).

A large-scale deep model pre-trained on massive labeled or unlabeled data transfers well to downstream tasks. Linear evaluation freezes parameters in the pre-trained model and trains a linear classifier separately, which is efficient and attractive for transfer. However, little work has investigated the classifier in linear evaluation except for the default logistic regression. Inspired by the statistical efficiency of naive Bayes, the paper revisits the classical topic on discriminative vs. generative classifiers. Theoretically, the paper considers the surrogate loss instead of the zero-one loss in analyses and generalizes the classical results from binary cases to multiclass ones. We show that, under mild assumptions, multiclass naive Bayes requires $O(\log n)$ samples to approach its asymptotic error while the corresponding multiclass logistic regression requires $O(n)$ samples, where $n$ is the feature dimension. To establish it, we present a multiclass $\mathcal{H}$-consistency bound framework and an explicit bound for logistic loss, which are of independent interests. Simulation results on a mixture of Gaussian validate our theoretical findings. Experiments on various pre-trained deep vision models show that naive Bayes consistently converges faster as the number of data increases. Besides, naive Bayes shows promise in few-shot cases and we observe the "two regimes” phenomenon in pre-trained supervised models. Our code is available at https://github.com/ML-GSAI/Revisiting-Dis-vs-Gen-Classifiers.

The non-local network (NLNet) presents a pioneering approach for capturing long-range dependencies within an image, via aggregating query-specific global context to each query position. However, through a rigorous empirical analysis, we have found that the global contexts modeled by the non-local network are almost the same for different query positions. In this paper, we take advantage of this finding to create a simplified network based on a query-independent formulation, which maintains the accuracy of NLNet but with significantly less computation. We further replace the one-layer transformation function of the non-local block by a two-layer bottleneck, which further reduces the parameter number considerably. The resulting network element, called the global context (GC) block, effectively models global context in a lightweight manner, allowing it to be applied at multiple layers of a backbone network to form a global context network (GCNet). Experiments show that GCNet generally outperforms NLNet on major benchmarks for various recognition tasks. The code and network configurations are available at https://github.com/xvjiarui/GCNet.

We present techniques for scaling Swin Transformer [35] up to 3 billion parameters and making it capable of training with images of up to 1,536x1,536 resolution. By scaling up capacity and resolution, Swin Transformer sets new records on four representative vision benchmarks: 84.0% top-1 accuracy on ImageNet- V2 image classification, 63.1 / 54.4 box / mask mAP on COCO object detection, 59.9 mIoU on ADE20K semantic segmentation, and 86.8% top-1 accuracy on Kinetics-400 video action classification. We tackle issues of training instability, and study how to effectively transfer models pre-trained at low resolutions to higher resolution ones. To this aim, several novel technologies are proposed: 1) a residual post normalization technique and a scaled cosine attention approach to improve the stability of large vision models; 2) a log-spaced continuous position bias technique to effectively transfer models pre-trained at low-resolution images and windows to their higher-resolution counterparts. In addition, we share our crucial implementation details that lead to significant savings of GPU memory consumption and thus make it feasi-ble to train large vision models with regular GPUs. Using these techniques and self-supervised pre-training, we suc-cessfully train a strong 3 billion Swin Transformer model and effectively transfer it to various vision tasks involving high-resolution images or windows, achieving the state-of-the-art accuracy on a variety of benchmarks. Code is avail-able at https://github.com/microsoft/Swin-Transformer.

The vision community is witnessing a modeling shift from CNNs to Transformers, where pure Transformer architectures have attained top accuracy on the major video recognition benchmarks. These video models are all built on Transformer layers that globally connect patches across the spatial and temporal dimensions. In this paper, we instead advocate an inductive bias of locality in video Transformers, which leads to a better speed-accuracy trade-off compared to previous approaches which compute self-attention globally even with spatial-temporal factorization. The locality of the proposed video architecture is realized by adapting the Swin Transformer designed for the image domain, while continuing to leverage the power of pre-trained image models. Our approach achieves state-of-the-art accuracy on a broad range of video recognition benchmarks, including on action recognition (84.9 top-l accuracy on Kinetics-400 and 85.9 top-l accuracy on Kinetics-600 with ~20× less pre-training data and ~3× smaller model size) and temporal modeling (69.6 top-l accuracy on Something-Something v2).

This paper presents SimMIM, a simple framework for masked image modeling. We have simplified recently proposed relevant approaches, without the need for special designs, such as block-wise masking and tokenization via discrete VAE or clustering. To investigate what makes a masked image modeling task learn good representations, we systematically study the major components in our framework, and find that the simple designs of each component have revealed very strong representation learning performance: 1) random masking of the input image with a moderately large masked patch size (e.g., 32) makes a powerful pre-text task; 2) predicting RGB values of raw pixels by direct regression performs no worse than the patch classification approaches with complex designs; 3) the prediction head can be as light as a linear layer, with no worse performance than heavier ones. Using ViT-B, our approach achieves 83.8% top-1 fine-tuning accuracy on ImageNet-1K by pre-training also on this dataset, surpassing previous best approach by +0.6%. When applied to a larger model with about 650 million parameters, SwinV2-H, it achieves 87.1% top-1 accuracy on ImageNet-1K using only ImageNet-1K data. We also leverage this approach to address the data-hungry issue faced by large-scale model training, that a 3B model (Swin V2-G) is successfully trained to achieve state-of-the-art accuracy on four representative vision benchmarks using 40× less labelled data than that in previous practice (JFT-3B). The code is available at https://github.com/microsoft/SimMIM.

Robustness and discrimination power are two fundamental requirements in visual object tracking. In most tracking paradigms, we find that the features extracted by the popular Siamese-like networks cannot fully discriminatively model the tracked targets and distractor objects, hindering them from simultaneously meeting these two requirements. While most methods focus on designing robust correlation operations, we propose a novel target-dependent feature network inspired by the self-/cross-attention scheme. In contrast to the Siamese-like feature extraction, our network deeply embeds cross-image feature correlation in multiple layers of the feature network. By extensively matching the features of the two images through multiple layers, it is able to suppress non-target features, resulting in instance-varying feature extraction. The output features of the search image can be directly used for predicting target locations without extra correlation step. Moreover, our model can be flexibly pre-trained on abundant unpaired images, leading to notably faster convergence than the existing methods. Extensive experiments show our method achieves the state-of-the-art results while running at real-time. Our feature networks also can be applied to existing tracking pipelines seamlessly to raise the tracking performance.

Frozen pretrained models have become a viable alternative to the pretraining-then-finetuning paradigm for transfer learning. However, with frozen models there are relatively few parameters available for adapting to downstream tasks, which is problematic in computer vision where tasks vary significantly in input/output format and the type of information that is of value. In this paper, we present a study of frozen pretrained models when applied to diverse and representative computer vision tasks, including object detection, semantic segmentation and video action recognition. From this empirical analysis, our work answers the questions of what pretraining task fits best with this frozen setting, how to make the frozen setting more flexible to various downstream tasks, and the effect of larger model sizes. We additionally examine the upper bound of performance using a giant frozen pretrained model with 3 billion parameters (SwinV2-G) and find that it reaches competitive performance on a varied set of major benchmarks with only one shared frozen base network: 60.0 box mAP and 52.2 mask mAP on COCO object detection test-dev, 57.6 val mIoU on ADE20K semantic segmentation, and 81.7 top-1 accuracy on Kinetics-400 action recognition. With this work, we hope to bring greater attention to this promising path of freezing pretrained image models.

Contrastive learning methods for unsupervised visual representation learning have reached remarkable levels of transfer performance. We argue that the power of contrastive learning has yet to be fully unleashed, as current methods are trained only on instance-level pretext tasks, leading to representations that may be sub-optimal for downstream tasks requiring dense pixel predictions. In this paper, we introduce pixel-level pretext tasks for learning dense feature representations. The first task directly applies contrastive learning at the pixel level. We additionally propose a pixel-to-propagation consistency task that produces better results, even surpassing the state-of-the-art approaches by a large margin. Specifically, it achieves 60.2 AP, 41.4 / 40.5 mAP and 77.2 mIoU when transferred to Pascal VOC object detection (C4), COCO object detection (FPN / C4) and Cityscapes semantic segmentation using a ResNet-50 backbone network, which are 2.6 AP, 0.8 / 1.0 mAP and 1.0 mIoU better than the previous best methods built on instance-level contrastive learning. Moreover, the pixel-level pretext tasks are found to be effective for pre-training not only regular backbone networks but also head networks used for dense downstream tasks, and are complementary to instance-level contrastive methods. These results demonstrate the strong potential of defining pretext tasks at the pixel level, and suggest a new path forward in unsupervised visual representation learning. Code is available at https://github.com/zdaxie/PixPro.

A well-known issue of Batch Normalization is its significantly reduced effectiveness in the case of small mini-batch sizes. When a mini-batch contains few examples, the statistics upon which the normalization is defined cannot be reliably estimated from it during a training iteration. To address this problem, we present Cross-Iteration Batch Normalization (CBN), in which examples from multiple recent iterations are jointly utilized to enhance estimation quality. A challenge of computing statistics over multiple iterations is that the network activations from different iterations are not comparable to each other due to changes in network weights. We thus compensate for the network weight changes via a proposed technique based on Taylor polynomials, so that the statistics can be accurately estimated and batch normalization can be effectively applied. On object detection and image classification with small mini-batch sizes, CBN is found to outperform the original batch normalization and a direct calculation of statistics over previous iterations without the proposed compensation technique. Code is available at https://aka.ms/cbn.

Recently, directly detecting 3D objects from 3D point clouds has received increasing attention. To extract object representation from an irregular point cloud, existing methods usually take a point grouping step to assign the points to an object candidate so that a PointNet-like network could be used to derive object features from the grouped points. However, the inaccurate point assignments caused by the hand-crafted grouping scheme decrease the performance of 3D object detection.In this paper, we present a simple yet effective method for directly detecting 3D objects from the 3D point cloud. Instead of grouping local points to each object candidate, our method computes the feature of an object from all the points in the point cloud with the help of an attention mechanism in the Transformers [42], where the contribution of each point is automatically learned in the network training. With an improved attention stacking scheme, our method fuses object features in different stages and generates more accurate object detection results. With few bells and whistles, the proposed method achieves state-of-the-art 3D object detection performance on two widely used benchmarks, Scan-Net V2 and SUN RGB-D. The code and models are publicly available at https://github.com/zeliu98/Group-Free-3D

This paper presents a new vision Transformer, called Swin Transformer, that capably serves as a general-purpose backbone for computer vision. Challenges in adapting Transformer from language to vision arise from differences between the two domains, such as large variations in the scale of visual entities and the high resolution of pixels in images compared to words in text. To address these differences, we propose a hierarchical Transformer whose representation is computed with Shifted windows. The shifted windowing scheme brings greater efficiency by limiting self-attention computation to non-overlapping local windows while also allowing for cross-window connection. This hierarchical architecture has the flexibility to model at various scales and has linear computational complexity with respect to image size. These qualities of Swin Transformer make it compatible with a broad range of vision tasks, including image classification (87.3 top-1 accuracy on ImageNet-1K) and dense prediction tasks such as object detection (58.7 box AP and 51.1 mask AP on COCO test-dev) and semantic segmentation (53.5 mIoU on ADE20K val). Its performance surpasses the previous state-of-the-art by a large margin of +2.7 box AP and +2.6 mask AP on COCO, and +3.2 mIoU on ADE20K, demonstrating the potential of Transformer-based models as vision backbones. The hierarchical design and the shifted window approach also prove beneficial for all-MLP architectures. The code and models are publicly available at https://github.com/microsoft/Swin-Transformer.

We introduce MixTraining, a new training paradigm for object detection that can improve the performance of existing detectors for free. MixTraining enhances data augmentation by utilizing augmentations of different strengths while excluding the strong augmentations of certain training samples that may be detrimental to training. In addition, it addresses localization noise and missing labels in human annotations by incorporating pseudo boxes that can compensate for these errors. Both of these MixTraining capabilities are made possible through bootstrapping on the detector, which can be used to predict the difficulty of training on a strong augmentation, as well as to generate reliable pseudo boxes thanks to the robustness of neural networks to labeling error. MixTraining is found to bring consistent improvements across various detectors on the COCO dataset. In particular, the performance of Faster R-CNN~\cite{ren2015faster} with a ResNet-50~\cite{he2016deep} backbone is improved from 41.7 mAP to 44.0 mAP, and the accuracy of Cascade-RCNN~\cite{cai2018cascade} with a Swin-Small~\cite{liu2021swin} backbone is raised from 50.9 mAP to 52.8 mAP.

How do humans recognize an object in a piece of video? Due to the deteriorated quality of single frame, it may be hard for people to identify an occluded object in this frame by just utilizing information within one image. We argue that there are two important cues for humans to recognize objects in videos: the global semantic information and the local localization information. Recently, plenty of methods adopt the self-attention mechanisms to enhance the features in key frame with either global semantic information or local localization information. In this paper we introduce memory enhanced global-local aggregation (MEGA) network, which is among the first trials that takes full consideration of both global and local information. Furthermore, empowered by a novel and carefully-designed Long Range Memory (LRM) module, our proposed MEGA could enable the key frame to get access to much more content than any previous methods. Enhanced by these two sources of information, our method achieves state-of-the-art performance on ImageNet VID dataset. Code is available at https://github.com/Scalsol/mega.pytorch.

This paper presents parametric instance classification (PIC) for unsupervised visual feature learning. Unlike the state-of-the-art approaches which do instance discrimination in a dual-branch non-parametric fashion, PIC directly performs a one-branch parametric instance classification, revealing a simple framework similar to supervised classification and without the need to address the information leakage issue. We show that the simple PIC framework can be as effective as the state-of-the-art approaches, i.e. SimCLR and MoCo v2, by adapting several common component settings used in the state-of-the-art approaches. We also propose two novel techniques to further improve effectiveness and practicality of PIC: 1) a sliding-window data scheduler, instead of the previous epoch-based data scheduler, which addresses the extremely infrequent instance visiting issue in PIC and improves the effectiveness; 2) a negative sampling and weight update correction approach to reduce the training time and GPU memory consumption, which also enables application of PIC to almost unlimited training images. We hope that the PIC framework can serve as a simple baseline to facilitate future study. The code and network configurations are available at \url{https://github.com/bl0/PIC}.

Verification and regression are two general methodologies for prediction in neural networks. Each has its own strengths: verification can be easier to infer accurately, and regression is more efficient and applicable to continuous target variables. Hence, it is often beneficial to carefully combine them to take advantage of their benefits. In this paper, we take this philosophy to improve state-of-the-art object detection, specifically by RepPoints. Though RepPoints provides high performance, we find that its heavy reliance on regression for object localization leaves room for improvement. We introduce verification tasks into the localization prediction of RepPoints, producing RepPoints v2, which proves consistent improvements of about 2.0 mAP over the original RepPoints on COCO object detection benchmark using different backbones and training methods. RepPoints v2 also achieves 52.1 mAP on the COCO \texttt{test-dev} by a single model. Moreover, we show that the proposed approach can more generally elevate other object detection frameworks as well as applications such as instance segmentation.

Deep hashing enables computation and memory efficient image search through end-to-end learning of feature representations and binary codes. While linear scan over binary hash codes is more efficient than over the high-dimensional representations, its linear-time complexity is still unacceptable for very large databases. Hamming space retrieval enables constant-time search through hash lookups, where for each query, there is a Hamming ball centered at the query and the data points within the ball are returned as relevant. Since inside the Hamming ball implies retrievable while outside irretrievable, it is crucial to explicitly characterize the Hamming ball. The main idea of this work is to directly embody the Hamming radius into the loss functions, leading to Maximum-Margin Hamming Hashing (MMHH), a new model specifically optimized for Hamming space retrieval. We introduce a max-margin t-distribution loss, where the t-distribution concentrates more similar data points to be within the Hamming ball, and the margin characterizes the Hamming radius such that less penalization is applied to similar data points within the Hamming ball. The loss function also introduces robustness to data noise, where the similarity supervision may be inaccurate in practical problems. The model is trained end-to-end using a new semi-batch optimization algorithm tailored to extremely imbalanced data. Our method yields state-of-the-art results on four datasets and shows superior performance on noisy data.

Recent progress in multiple object tracking (MOT) has shown that a robust similarity score is a key to the success of trackers. A good similarity score is expected to reflect multiple cues, e.g. appearance, location, and topology, over a long period of time. However, these cues are heterogeneous, making them hard to be combined in a unified network. As a result, existing methods usually encode them in separate networks or require a complex training approach. In this paper, we present a unified framework for similarity measurement based on spatial-temporal relation network which could simultaneously encode various cues and perform reasoning across both spatial and temporal domains. We also study the feature representation of a tracklet-object pair in depth, showing a proper design of the pair features can well empower the trackers. The resulting approach is named spatial-temporal relation networks (STRN). It runs in a feed-forward way and can be trained in an end-to-end manner. The state-of-the-art accuracy was achieved on all of the MOT15$\sim$17 benchmarks using public detection and online settings.

Domain adaptation studies learning algorithms that generalize across source domains and target domains that exhibit different distributions. Recent studies reveal that deep neural networks can learn transferable features that generalize well to similar novel tasks. However, as deep features eventually transition from general to specific along the network, feature transferability drops significantly in higher task-specific layers with increasing domain discrepancy. To formally reduce the effects of this discrepancy and enhance feature transferability in task-specific layers, we develop a novel framework for deep adaptation networks that extends deep convolutional neural networks to domain adaptation problems. The framework embeds the deep features of all task-specific layers into reproducing kernel Hilbert spaces (RKHSs) and optimally matches different domain distributions. The deep features are made more transferable by exploiting low-density separation of target-unlabeled data in very deep architectures, while the domain discrepancy is further reduced via the use of multiple kernel learning that enhances the statistical power of kernel embedding matching. The overall framework is cast in a minimax game setting. Extensive empirical evidence shows that the proposed networks yield state-of-the-art results on standard visual domain-adaptation benchmarks.

Domain adaptation generalizes a learning model across source domain and target domain that follow different distributions. Most existing work follows a two-step procedure: first, explores either feature matching or instance reweighting independently, and second, train the transfer classifier separately. In this paper, we show that either feature matching or instance reweighting can only reduce, but not remove, the cross-domain discrepancy, and the knowledge hidden in the relations between the data labels from the source and target domains is important for unsupervised domain adaptation. We propose a new Distribution Matching Machine (DMM) based on the structural risk minimization principle, which learns a transfer support vector machine by extracting invariant feature representations and estimating unbiased instance weights that jointly minimize the cross-domain distribution discrepancy. This leads to a robust transfer learner that performs well against both mismatched features and irrelevant instances. Our theoretical analysis proves that the proposed approach further reduces the generalization error bound of related domain adaptation methods. Comprehensive experiments validate that the DMM approach significantly outperforms competitive methods on standard domain adaptation benchmarks.

Due to its computation efficiency and retrieval quality, hashing has been widely applied to approximate nearest neighbor search for large-scale image retrieval, while deep hashing further improves the retrieval quality by end-to-end representation learning and hash coding. With compact hash codes, Hamming space retrieval enables the most efficient constant-time search that returns data points within a given Hamming radius to each query, by hash table lookups instead of linear scan. However, subject to the weak capability of concentrating relevant images to be within a small Hamming ball due to mis-specified loss functions, existing deep hashing methods may underperform for Hamming space retrieval. This work presents Deep Cauchy Hashing (DCH), a novel deep hashing model that generates compact and concentrated binary hash codes to enable efficient and effective Hamming space retrieval. The main idea is to design a pairwise cross-entropy loss based on Cauchy distribution, which penalizes significantly on similar image pairs with Hamming distance larger than the given Hamming radius threshold. Comprehensive experiments demonstrate that DCH can generate highly concentrated hash codes and yield state-of-the-art Hamming space retrieval performance on three datasets, NUS-WIDE, CIFAR-10, and MS-COCO.

Deep learning to hash improves image retrieval performance by end-to-end representation learning and hash coding from training data with pairwise similarity information. Subject to the scarcity of similarity information that is often expensive to collect for many application domains, existing deep learning to hash methods may overfit the training data and result in substantial loss of retrieval quality. This paper presents HashGAN, a novel architecture for deep learning to hash, which learns compact binary hash codes from both real images and diverse images synthesized by generative models. The main idea is to augment the training data with nearly real images synthesized from a new Pair Conditional Wasserstein GAN (PC-WGAN) conditioned on the pairwise similarity information. Extensive experiments demonstrate that HashGAN can generate high-quality binary hash codes and yield state-of-the-art image retrieval performance on three benchmarks, NUS-WIDE, CIFAR-10, and MS-COCO.

Deep hashing establishes efficient and effective image retrieval by end-to-end learning of deep representations and hash codes from similarity data. We present a compact coding solution, focusing on deep learning to quantization approach that has shown superior performance over hashing solutions for similarity retrieval. We propose Deep Triplet Quantization (DTQ), a novel approach to learning deep quantization models from the similarity triplets. To enable more effective triplet training, we design a new triplet selection approach, Group Hard, that randomly selects hard triplets in each image group. To generate compact binary codes, we further apply a triplet quantization with weak orthogonality during triplet training. The quantization loss reduces the codebook redundancy and enhances the quantizability of deep representations through back-propagation. Extensive experiments demonstrate that DTQ can generate high-quality and compact binary codes, which yields state-of-the-art image retrieval performance on three benchmark datasets, NUS-WIDE, CIFAR-10, and MS-COCO.

Cross-modal similarity retrieval is a problem about designing a retrieval system that supports querying across content modalities, e.g., using an image to retrieve for texts. This paper presents a compact coding solution for efficient cross-modal retrieval, with a focus on the quantization approach which has already shown the superior performance over the hashing solutions in single-modal similarity retrieval. We propose a collective deep quantization (CDQ) approach, which is the first attempt to introduce quantization in end-to-end deep architecture for cross-modal retrieval. The major contribution lies in jointly learning deep representations and the quantizers for both modalities using carefully-crafted hybrid networks and well-specified loss functions. In addition, our approach simultaneously learns the common quantizer codebook for both modalities through which the cross-modal correlation can be substantially enhanced. CDQ enables efficient and effective cross-modal retrieval using inner product distance computed based on the common codebook with fast distance table lookup. Extensive experiments show that CDQ yields state of the art cross-modal retrieval results on standard benchmarks.

Compact coding has been widely applied to approximate nearest neighbor search for large-scale image retrieval, due to its computation efficiency and retrieval quality. This paper presents a compact coding solution with a focus on the deep learning to quantization approach, which improves retrieval quality by end-to-end representation learning and compact encoding and has already shown the superior performance over the hashing solutions for similarity retrieval. We propose Deep Visual-Semantic Quantization (DVSQ), which is the first approach to learning deep quantization models from labeled image data as well as the semantic information underlying general text domains. The main contribution lies in jointly learning deep visual-semantic embeddings and visual-semantic quantizers using carefully-designed hybrid networks and well-specified loss functions. DVSQ enables efficient and effective image retrieval by supporting maximum inner-product search, which is computed based on learned codebooks with fast distance table lookup. Comprehensive empirical evidence shows that DVSQ can generate compact binary codes and yield state-of-the-art similarity retrieval performance on standard benchmarks.

Hashing has been widely applied to approximate nearest neighbor search for large-scale multimedia retrieval. Supervised hashing improves the quality of hash coding by exploiting the semantic similarity on data pairs and has received increasing attention recently. For most existing supervised hashing methods for image retrieval, an image is first represented as a vector of hand-crafted or machine-learned features, then quantized by a separate quantization step that generates binary codes. However, suboptimal hash coding may be produced, since the quantization error is not statistically minimized and the feature representation is not optimally compatible with the hash coding. In this paper, we propose a novel Deep Quantization Network (DQN) architecture for supervised hashing, which learns image representation for hash coding and formally control the quantization error. The DQN model constitutes four key components: (1) a sub-network with multiple convolution-pooling layers to capture deep image representations; (2) a fully connected bottleneck layer to generate dimension-reduced representation optimal for hash coding; (3) a pairwise cosine loss layer for similarity-preserving learning; and (4) a product quantization loss for controlling hashing quality and the quantizability of bottleneck representation. Extensive experiments on standard image retrieval datasets show the proposed DQN model yields substantial boosts over latest state-of-the-art hashing methods.

Due to the storage and retrieval efficiency, hashing has been widely deployed to approximate nearest neighbor search for large-scale multimedia retrieval. Supervised hashing, which improves the quality of hash coding by exploiting the semantic similarity on data pairs, has received increasing attention recently. For most existing supervised hashing methods for image retrieval, an image is first represented as a vector of hand-crafted or machine-learned features, followed by another separate quantization step that generates binary codes. However, suboptimal hash coding may be produced, because the quantization error is not statistically minimized and the feature representation is not optimally compatible with the binary coding. In this paper, we propose a novel Deep Hashing Network (DHN) architecture for supervised hashing, in which we jointly learn good image representation tailored to hash coding and formally control the quantization error. The DHN model constitutes four key components: (1) a sub-network with multiple convolution-pooling layers to capture image representations; (2) a fully-connected hashing layer to generate compact binary hash codes; (3) a pairwise cross-entropy loss layer for similarity-preserving learning; and (4) a pairwise quantization loss for controlling hashing quality. Extensive experiments on standard image retrieval datasets show the proposed DHN model yields substantial boosts over latest state-of-the-art hashing methods.

Due to the storage and retrieval efficiency, hashing has been widely applied to approximate nearest neighbor search for large-scale multimedia retrieval. Cross-modal hashing, which enables efficient retrieval of images in response to text queries or vice versa, has received increasing attention recently. Most existing work on cross-modal hashing does not capture the spatial dependency of images and temporal dynamics of text sentences for learning powerful feature representations and cross-modal embeddings that mitigate the heterogeneity of different modalities. This paper presents a new Deep Visual-Semantic Hashing (DVSH) model that generates compact hash codes of images and sentences in an end-to-end deep learning architecture, which capture the intrinsic cross-modal correspondences between visual data and natural language. DVSH is a hybrid deep architecture that constitutes a visual-semantic fusion network for learning joint embedding space of images and text sentences, and two modality-specific hashing networks for learning hash functions to generate compact binary codes. Our architecture effectively unifies joint multimodal embedding and cross-modal hashing, which is based on a novel combination of Convolutional Neural Networks over images, Recurrent Neural Networks over sentences, and a structured max-margin objective that integrates all things together to enable learning of similarity-preserving and high-quality hash codes. Extensive empirical evidence shows that our DVSH approach yields state of the art results in cross-modal retrieval experiments on image-sentences datasets, i.e. standard IAPR TC-12 and large-scale Microsoft COCO.

Due to its storage and query efficiency, hashing has been widely applied to approximate nearest neighbor search from large-scale datasets. While there is increasing interest in cross-modal hashing which facilitates cross-media retrieval by embedding data from different modalities into a common Hamming space, how to distill the cross-modal correlation structure effectively remains a challenging problem. In this paper, we propose a novel supervised cross-modal hashing method, Correlation Autoencoder Hashing (CAH), to learn discriminative and compact binary codes based on deep autoencoders. Specifically, CAH jointly maximizes the feature correlation revealed by bimodal data and the semantic correlation conveyed in similarity labels, while embeds them into hash codes by nonlinear deep autoencoders. Extensive experiments clearly show the superior effectiveness and efficiency of CAH against the state-of-the-art hashing methods on standard cross-modal retrieval benchmarks.

Efficient similarity retrieval from large-scale multimodal database is pervasive in modern search engines and social networks. To support queries across content modalities, the system should enable cross-modal correlation and computation-efficient indexing. While hashing methods have shown great potential in achieving this goal, current attempts generally fail to learn isomorphic hash codes in a seamless scheme, that is, they embed multiple modalities in a continuous isomorphic space and separately threshold embeddings into binary codes, which incurs substantial loss of retrieval accuracy. In this paper, we approach seamless multimodal hashing by proposing a novel Composite Correlation Quantization (CCQ) model. Specifically, CCQ jointly finds correlation-maximal mappings that transform different modalities into isomorphic latent space, and learns composite quantizers that convert the isomorphic latent features into compact binary codes. An optimization framework is devised to preserve both intra-modal similarity and inter-modal correlation through minimizing both reconstruction and quantization errors, which can be trained from both paired and partially paired data in linear time. A comprehensive set of experiments clearly show the superior effectiveness and efficiency of CCQ against the state of the art hashing methods for both unimodal and cross-modal retrieval.

Timestamps are often found to be dirty in various scenarios, e.g., in distributed systems with clock synchronization problems or unreliable RFID readers. Without cleaning the imprecise timestamps, temporal-related applications such as provenance analysis or pattern queries are not reliable. To evaluate the correctness of timestamps, temporal constraints could be employed, which declare the distance restrictions between timestamps. Guided by such constraints on timestamps, in this paper, we study a novel problem of repairing inconsistent timestamps that do not conform to the required temporal constraints. Following the same line of data repairing, the timestamp repairing problem is to minimally modify the timestamps towards satisfaction of temporal constraints. This problem is practically challenging, given the huge space of possible timestamps. We tackle the problem by identifying a concise set of promising candidates, where an optimal repair solution can always be found. Repair algorithms with efficient pruning are then devised over the identified candidates. Experiments on real datasets demonstrate the superiority of our proposal compared to the state-of-the-art approaches.

Domain adaptation generalizes a learning model across source domain and target domain that are sampled from different distributions. It is widely applied to cross-domain data mining for reusing labeled information and mitigating labeling consumption. Recent studies reveal that deep neural networks can learn abstract feature representation, which can reduce, but not remove, the cross-domain discrepancy. To enhance the invariance of deep representation and make it more transferable across domains, we propose a unified deep adaptation framework for jointly learning transferable representation and classifier to enable scalable domain adaptation, by taking the advantages of both deep learning and optimal two-sample matching. The framework constitutes two inter-dependent paradigms, unsupervised pre-training for effective training of deep models using deep denoising autoencoders, and supervised fine-tuning for effective exploitation of discriminative information using deep neural networks, both learned by embedding the deep representations to reproducing kernel Hilbert spaces (RKHSs) and optimally matching different domain distributions. To enable scalable learning, we develop a linear-time algorithm using unbiased estimate that scales linearly to large samples. Extensive empirical results show that the proposed framework significantly outperforms state of the art methods on diverse adaptation tasks: sentiment polarity prediction, email spam filtering, newsgroup content categorization, and visual object recognition.

Recent studies reveal that a deep neural network can learn transferable features which generalize well to novel tasks for domain adaptation. However, as deep features eventually transition from general to specific along the network, the feature transferability drops significantly in higher layers with increasing domain discrepancy. Hence, it is important to formally reduce the dataset bias and enhance the transferability in task-specific layers. In this paper, we propose a new Deep Adaptation Network (DAN) architecture, which generalizes deep convolutional neural network to the domain adaptation scenario. In DAN, hidden representations of all task-specific layers are embedded in a reproducing kernel Hilbert space where the mean embeddings of different domain distributions can be explicitly matched. The domain discrepancy is further reduced using an optimal multi-kernel selection method for mean embedding matching. DAN can learn transferable features with statistical guarantees, and can scale linearly by unbiased estimate of kernel embedding. Extensive empirical evidence shows that the proposed architecture yields state-of-the-art image classification error rates on standard domain adaptation benchmarks.