Recent advances in image synthesis have been propelled by powerful generative models, such as Masked Generative Transformers (MaskGIT), autoregressive models, diffusion models, and rectified flow models. A common principle behind their success is the decomposition of complex synthesis tasks into multiple tractable steps. However, this introduces a proliferation of step-specific parameters to be configured for modulating the iterative generation process (e.g., mask ratio, noise level, or temperature at each step). Existing approaches typically rely on manually-designed scheduling rules to manage this complexity, demanding expert knowledge and extensive trial-and-error. Furthermore, these static schedules lack the flexibility to adapt to the unique characteristics of each individual sample, yielding sub-optimal performance. To address this issue, we present AdaGen, a general, learnable, and sample-adaptive framework for scheduling the iterative generation process. Specifically, we formulate the scheduling problem as a Markov Decision Process, where a lightweight policy network is introduced to adaptively determine the most suitable parameters given the current generation state, and can be trained through reinforcement learning. Importantly, we demonstrate that simple reward designs, such as FID or pre-trained reward models, can be easily hacked and may not reliably guarantee the desired quality or diversity of generated samples. Therefore, we propose an adversarial reward design to guide the training of the policy networks effectively. Finally, we introduce an inference-time refinement strategy and a controllable fidelity-diversity trade-off mechanism to further enhance the performance and flexibility of AdaGen. Comprehensive experiments across five benchmark datasets (ImageNet-256 × 256 & 512 × 512, MS-COCO, CC3M, and LAION-5B) and four distinct generative paradigms validate the superiority of AdaGen . For example, AdaGen achieves better performance on DiT-XL with $\mathbf {\sim 3\times }$∼3× lower inference cost and improves the FID of VAR from 1.92 to 1.59 with negligible additional computational overhead.

The astonishing breakthrough of multimodal large language models (MLLMs) has necessitated new benchmarks to quantitatively assess their capabilities, reveal their limitations, and indicate future research directions. However, this is challenging in the context of remote sensing (RS), since the imagery features ultra-high resolution that incorporates extremely complex semantic relationships. Existing benchmarks usually adopt notably smaller image sizes than real-world RS scenarios, suffer from limited annotation quality, and consider insufficient dimensions of evaluation. To address these issues, we present XLRS-Bench: a comprehensive benchmark for evaluating the perception and reasoning capabilities of MLLMs in ultra-high-resolution RS scenarios. XLRS-Bench boasts the largest average image size (8500×8500) observed thus far, with all evaluation samples meticulously annotated manually, assisted by a novel semi-automatic captioner on ultra-high-resolution RS images. On top of the XLRS-Bench, 16 sub-tasks are defined to evaluate MLLMs’ 10 kinds of perceptual capabilities and 6 kinds of reasoning capabilities, with a primary emphasis on advanced cognitive processes that facilitate real-world decision-making and the capture of spatiotemporal changes. The results of both general and RS-focused MLLMs on XLRS-Bench indicate that further efforts are needed for real-world RS applications. We have open-sourced XLRS-Bench to support further research in developing more powerful MLLMs for remote sensing.

Test-Time adaptation (TTA) aims to improve the performance of source-domain pre-trained models on previously unseen, shifted target domains. Traditional TTA methods primarily adapt model weights based on target data streams, making model performance sensitive to the amount and order of target data. The recently proposed diffusion-driven TTA methods mitigate this by adapting model inputs instead of weights, where an unconditional diffusion model, trained on the source domain, transforms target-domain data into a synthetic domain that is expected to approximate the source domain. However, in this paper, we reveal that although the synthetic data in diffusion-driven TTA seems indistinguishable from the source data, it is unaligned with, or even markedly different from the latter for deep networks. To address this issue, we propose a Synthetic-Domain Alignment (SDA) framework. Our key insight is to fine-tune the source model with synthetic data to ensure better alignment. Specifically, we first employ a conditional diffusion model to generate labeled samples, creating a synthetic dataset. Subsequently, we use the aforementioned unconditional diffusion model to add noise to and denoise each sample before fine-tuning. This Mix of Diffusion (MoD) process mitigates the potential domain misalignment between the conditional and unconditional models. Extensive experiments across classifiers, segmenters, and multimodal large language models (MLLMs, e.g., LLaVA) demonstrate that SDA achieves superior domain alignment and consistently outperforms existing diffusion-driven TTA methods. Our code is available at https://github.com/SHI-Labs/Diffusion-Driven-Test-Time-Adaptation-Via-Synthetic-Domain-Alignment.

Echocardiography is crucial for cardiovascular disease detection but relies heavily on experienced sonographers. Echocardiography probe guidance systems, which provide real-time movement instructions for acquiring standard plane images, offer a promising solution for AI-assisted or fully autonomous scanning. However, developing effective machine learning models for this task remains challenging, as they must grasp heart anatomy and the intricate interplay between probe motion and visual signals. To address this, we present EchoWorld, a motion-aware world modeling framework for probe guidance that encodes anatomical knowledge and motion-induced visual dynamics, while effectively leveraging past visual-motion sequences to enhance guidance precision. EchoWorld employs a pre-training strategy inspired by world modeling principles, where the model predicts masked anatomical regions and simulates the visual outcomes of probe adjustments. Built upon this pre-trained model, we introduce a motion-aware attention mechanism in the fine-tuning stage that effectively integrates historical visual-motion data, enabling precise and adaptive probe guidance. Trained on more than one million ultrasound images from over 200 routine scans, EchoWorld effectively captures key echocar-diographic knowledge, as validated by qualitative analysis. Moreover, our method significantly reduces guidance errors compared to existing visual backbones and guidance frameworks, excelling in both single-frame and sequential evaluation protocols. Code is available at https://github.com/LeapLabTHU/EchoWorld.

Humans can develop internal world models that encode common sense knowledge, telling them how the world works and predicting the consequences of their actions. This concept has emerged as a promising direction for establishing general-purpose machine-learning models in recent preliminary works, e.g., for visual representation learning. In this paper, we present CheXWorld, the first effort towards a self-supervised world model for radiographic images. Specifically, our work develops a unified framework that simultaneously models three aspects of medical knowledge essential for qualified radiologists, including 1) local anatomical structures describing the fine-grained characteristics of local tissues (e.g., architectures, shapes, and textures); 2) global anatomical layouts describing the global organization of the human body (e.g., layouts of organs and skeletons); and 3) domain variations that encourage CheXWorld to model the transitions across different appearance domains of radiographs (e.g., varying clarity, contrast, and exposure caused by collecting radiographs from different hospitals, devices, or patients). Empirically, we design tailored qualitative and quantitative analyses, revealing that CheX-World successfully captures these three dimensions of medical knowledge. Furthermore, transfer learning experiments across eight medical image classification and segmentation benchmarks showcase that CheXWorld significantly outperforms existing SSL methods and large-scale medical foundation models. Code & pre-trained models are available at https://github.com/LeapLabTHU/CheXWorld.

Ultra-high-resolution (UHR) remote sensing (RS) imagery offers valuable data for Earth observation but pose challenges for existing multimodal foundation models due to two key bottlenecks: (1) limited availability of UHR training data, and (2) token explosion caused by the large image size. To address data scarcity, we introduce **SuperRS-VQA** (avg. 8,376$\times$8,376) and **HighRS-VQA** (avg. 2,000$\times$1,912), the highest-resolution vision-language datasets in RS to date, covering 22 real-world dialogue tasks. To mitigate token explosion, our pilot studies reveal significant redundancy in RS images: crucial information is concentrated in a small subset of object-centric tokens, while pruning background tokens (e.g., ocean or forest) can even improve performance. Motivated by these findings, we propose two strategies: *Background Token Pruning* and *Anchored Token Selection*, to reduce the memory footprint while preserving key semantics. Integrating these techniques, we introduce **GeoLLaVA-8K**, the first RS-focused multimodal large language model capable of handling inputs up to 8K$\times$8K resolution, built on the LLaVA framework. Trained on SuperRS-VQA and HighRS-VQA, GeoLLaVA-8K sets a new state-of-the-art on the XLRS-Bench. Datasets and code were released at https://github.com/MiliLab/GeoLLaVA-8K.

Recent advances in self-supervised learning for Vision Transformers (ViTs) have fueled breakthroughs in remote sensing (RS) foundation models. However, the quadratic complexity of self-attention poses a significant barrier to scalability, particularly for large models and high-resolution images. While the linear-complexity Mamba architecture offers a promising alternative, existing RS applications of Mamba remain limited to supervised tasks on small, domain-specific datasets. To address these challenges, we propose RoMA, a framework that enables scalable self-supervised pretraining of Mamba-based RS foundation models using large-scale, diverse, unlabeled data. RoMA enhances scalability for high-resolution images through a tailored auto-regressive learning strategy, incorporating two key innovations: 1) a rotation-aware pretraining mechanism combining adaptive cropping with angular embeddings to handle sparsely distributed objects with arbitrary orientations, and 2) multi-scale token prediction objectives that address the extreme variations in object scales inherent to RS imagery. Systematic empirical studies validate that Mamba adheres to RS data and parameter scaling laws, with performance scaling reliably as model and data size increase. Furthermore, experiments across scene classification, object detection, and semantic segmentation tasks demonstrate that RoMA-pretrained Mamba models consistently outperform ViT-based counterparts in both accuracy and computational efficiency. The source code and pretrained models have be released at https://github.com/MiliLab/RoMA.

This paper presents a comprehensive exploration of the phenomenon of data redundancy in video understanding, with the aim to improve computational efficiency. Our investigation commences with an examination of spatial redundancy, which refers to the observation that the most informative region in each video frame usually corresponds to a small image patch, whose shape, size and location shift smoothly across frames. Motivated by this phenomenon, we formulate the patch localization problem as a dynamic decision task, and introduce a spatially adaptive video recognition approach, termed AdaFocus. In specific, a lightweight encoder is first employed to quickly process the full video sequence, whose features are then utilized by a policy network to identify the most task-relevant regions. Subsequently, the selected patches are inferred by a high-capacity deep network for the final prediction. The complete model can be trained conveniently in an end-to-end manner. During inference, once the informative patch sequence has been generated, the bulk of computation can be executed in parallel, rendering it efficient on modern GPU devices. Furthermore, we demonstrate that AdaFocus can be easily extended by further considering the temporal and sample- wise redundancies, i.e., allocating the majority of computation to the most task-relevant video frames, and minimizing the computation spent on relatively "easier" videos. Our resulting algorithm, Uni-AdaFocus, establishes a comprehensive framework that seamlessly integrates spatial, temporal, and sample- wise dynamic computation, while it preserves the merits of AdaFocus in terms of efficient end-to-end training and hardware friendliness. In addition, Uni-AdaFocus is general and flexible as it is compatible with off-the-shelf backbone models (e.g., TSM and X3D), which can be readily deployed as our feature extractor, yielding a significantly improved computational efficiency. Empirically, extensive experiments based on seven widely-used benchmark datasets (i.e., ActivityNet, FCVID, Mini-Kinetics, Something-Something V1&V2, Jester, and Kinetics-400) and three real-world application scenarios (i.e., fine-grained diving action classification, Alzheimer's and Parkinson's diseases diagnosis with brain magnetic resonance images (MRI), and violence recognition for online videos) substantiate that Uni-AdaFocus is considerably more efficient than the competitive baselines. Code and pre-trained models are available at https://github.com/blackfeather-wang/AdaFocus and https://github.com/LeapLabTHU/AdaFocusV2.

The field of image synthesis is currently flourishing due to the advancements in diffusion models. While diffusion models have been successful, their computational inten-sity has prompted the pursuit of more efficient alternatives. As a representative work, non-autoregressive Transformers (NATs) have been recognized for their rapid generation. However, a major drawback of these models is their in-ferior performance compared to diffusion models. In this paper, we aim to re-evaluate the full potential of NATs by revisiting the design of their training and inference strategies. Specifically, we identify the complexities in properly configuring these strategies and indicate the possible sub-optimality in existing heuristic-driven designs. Recognizing this, we propose to go beyond existing methods by directly solving the optimal strategies in an automatic framework. The resulting method, named AutoNAT, advances the performance boundaries of NATs notably, and is able to perform comparably with the latest diffusion models with a significantly reduced inference cost. The effectiveness of AutoNAT is comprehensively validated on four benchmark datasets, i.e., ImageNet-256 & 512, MS-COCO, and CC3M. Code and pretrained models will be available at htt P s: / /gi thub. com/LeapLabTHU/ImprovedNAT.

Recently, token-based generation approaches have demonstrated their effectiveness in synthesizing visual content. As a representative example, non-autoregressive Transformers (NATs) can generate decent-quality images in just a few steps. NATs perform generation in a progressive manner, where the latent tokens of a resulting image are incrementally revealed step-by-step. At each step, the unrevealed image regions are padded with [MASK] tokens and inferred by NAT, with the most reliable predictions preserved as newly revealed, visible tokens. In this paper, we delve into understanding the mechanisms behind the effectiveness of NATs and uncover two important interaction patterns that naturally emerge from NAT’s paradigm: Spatially (within a step), although [MASK] and visible tokens are processed uniformly by NATs, the interactions between them are highly asymmetric. In specific, [MASK] tokens mainly gather information for decoding. On the contrary, visible tokens tend to primarily provide information, and their deep representations can be built only upon themselves. Temporally (across steps), the interactions between adjacent generation steps mostly concentrate on updating the representations of a few critical tokens, while the computation for the majority of tokens is generally repetitive. Driven by these findings, we propose EfficientNAT (ENAT), a NAT model that explicitly encourages these critical interactions inherent in NATs. At the spatial level, we disentangle the computations of visible and [MASK] tokens by encoding visible tokens independently, while decoding [MASK] tokens conditioned on the fully encoded visible tokens. At the temporal level, we prioritize the computation of the critical tokens at each step, while maximally reusing previously computed token representations to supplement necessary information. ENAT improves the performance of NATs notably with significantly reduced computational cost. Experiments on ImageNet-256 2 & 512 2 and MS-COCO validate the effectiveness of ENAT. Code and pre-trained models will be released at https://github.com/LeapLabTHU/ENAT.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable comprehension and reasoning capabilities with complex language and visual data. These advances have spurred the vision of establishing a generalist robotic MLLM proficient in understanding complex human instructions and accomplishing various embodied tasks, whose feasibility has been recently verified~\cite{rt-2,rt-x}. However, developing MLLMs for real-world robots is challenging due to the typically limited computation and memory capacities available on robotic platforms. In contrast, the inference of MLLMs usually incorporates storing billions of parameters and performing tremendous computation, imposing significant hardware demands. In our paper, we seek to address this challenge by leveraging an intriguing observation: relatively easier situations make up the bulk of the procedure of controlling robots to fulfill diverse tasks, and they generally require far smaller models to obtain the correct robotic actions. Motivated by this observation, we propose a \emph{Dynamic Early-Exit for Robotic MLLM} (DeeR) framework that automatically adjusts the size of the activated MLLM based on each situation at hand. The approach leverages a multi-exit architecture in MLLMs, which allows the model to cease processing once a proper size of the model has been activated for a specific situation, thus avoiding further redundant computation. Additionally, we develop novel algorithms that establish early-termination criteria for DeeR, conditioned on predefined demands such as average computational cost (\emph{i.e.}, power consumption), as well as peak computational consumption (\emph{i.e.}, latency) and GPU memory usage. These enhancements ensure that DeeR operates efficiently under varying resource constraints while maintaining competitive performance. Moreover, we design a tailored training method for integrating temporal information on top of such multi-exit architectures to predict actions reasonably. On the CALVIN robot manipulation benchmark, DeeR demonstrates significant reductions in computational costs by 5.2-6.5x and GPU memory by 2x without compromising performance. Code and checkpoints are available at https://github.com/yueyang130/DeeR-VLA.

Long-tailed distributions frequently emerge in real-world data, where a large number of minority categories contain a limited number of samples. Such imbalance issue considerably impairs the performance of standard supervised learning algorithms, which are mainly designed for balanced training sets. Recent investigations have revealed that supervised contrastive learning exhibits promising potential in alleviating the data imbalance. However, the performance of supervised contrastive learning is plagued by an inherent challenge: it necessitates sufficiently large batches of training data to construct contrastive pairs that cover all categories, yet this requirement is difficult to meet in the context of class-imbalanced data. To overcome this obstacle, we propose a novel probabilistic contrastive (ProCo) learning algorithm that estimates the data distribution of the samples from each class in the feature space, and samples contrastive pairs accordingly. In fact, estimating the distributions of all classes using features in a small batch, particularly for imbalanced data, is not feasible. Our key idea is to introduce a reasonable and simple assumption that the normalized features in contrastive learning follow a mixture of von Mises-Fisher (vMF) distributions on unit space, which brings two-fold benefits. First, the distribution parameters can be estimated using only the first sample moment, which can be efficiently computed in an online manner across different batches. Second, based on the estimated distribution, the vMF distribution allows us to sample an infinite number of contrastive pairs and derive a closed form of the expected contrastive loss for efficient optimization. Other than long-tailed problems, ProCo can be directly applied to semi-supervised learning by generating pseudo-labels for unlabeled data, which can subsequently be utilized to estimate the distribution of the samples inversely. Theoretically, we analyze the error bound of ProCo. Empirically, extensive experimental results on supervised/semi-supervised visual recognition and object detection tasks demonstrate that ProCo consistently outperforms existing methods across various datasets.

The superior performance of modern computer vision backbones (e.g., vision Transformers learned on ImageNet-1 K/22 K) usually comes with a costly training procedure. This study contributes to this issue by generalizing the idea of curriculum learning beyond its original formulation, i.e., training models using easier-to-harder data. Specifically, we reformulate the training curriculum as a soft-selection function, which uncovers progressively more difficult patterns within each example during training, instead of performing easier-to-harder sample selection. Our work is inspired by an intriguing observation on the learning dynamics of visual backbones: during the earlier stages of training, the model predominantly learns to recognize some 'easier-to-learn' discriminative patterns in the data. These patterns, when observed through frequency and spatial domains, incorporate lower-frequency components, and the natural image contents without distortion or data augmentation. Motivated by these findings, we propose a curriculum where the model always leverages all the training data at every learning stage, yet the exposure to the 'easier-to-learn' patterns of each example is initiated first, with harder patterns gradually introduced as training progresses. To implement this idea in a computationally efficient way, we introduce a cropping operation in the Fourier spectrum of the inputs, enabling the model to learn from only the lower-frequency components. Then we show that exposing the contents of natural images can be readily achieved by modulating the intensity of data augmentation. Finally, we integrate these two aspects and design curriculum learning schedules by proposing tailored searching algorithms. Moreover, we present useful techniques for deploying our approach efficiently in challenging practical scenarios, such as large-scale parallel training, and limited input/output or data pre-processing speed. The resulting method, EfficientTrain++, is simple, general, yet surprisingly effective. As an off-the-shelf approach, it reduces the training time of various popular models (e.g., ResNet, ConvNeXt, DeiT, PVT, Swin, CSWin, and CAFormer) by [Formula: see text] on ImageNet-1 K/22 K without sacrificing accuracy. It also demonstrates efficacy in self-supervised learning (e.g., MAE).

Recent years have witnessed a growing interest in compressed video action recognition due to the rapid growth of online videos. It remarkably reduces the storage by replacing raw videos with sparsely sampled RGB frames and other compressed motion cues (motion vectors and residuals). However, existing compressed video action recognition methods face two main issues: First, the inefficiency caused by the usage of coarse-level information under full resolution, and second, the disturbing due to the noisy dynamics in motion vectors. To address the two issues, this paper proposes a dynamic spatial focus method for efficient compressed video action recognition (CoViFocus). Specifically, we first use a light-weighted two-stream architecture to localize the task-relevant patches for both the RGB frames and motion vectors. Then the selected patch pair will be processed by a high-capacity two-stream deep model for the final prediction. Such a patch selection strategy crops out the irrelevant motion noise in motion vectors, as well as reduces the spatial redundancy of the inputs, leading to the high efficiency of our method in the compressed domain. Moreover, we found that the motion vectors can help our method to address the possibly happened static-issue, which means that the focus patches get stuck at some regions related to static objects rather than target actions, which further improves our method. Extensive results on both the HMDB-51 and UCF-101 datasets demonstrate the effectiveness and efficiency of our method in compressed video action recognition tasks.

Fine-grained image recognition is a longstanding computer vision challenge that focuses on differentiating objects belonging to multiple subordinate categories within the same meta-category. Since images belonging to the same meta-category usually share similar visual appearances, mining discriminative visual cues is the key to distinguishing fine-grained categories. Although commonly used image-level data augmentation techniques have achieved great success in generic image classification problems, they are rarely applied in fine-grained scenarios, because their random editing-region behavior is prone to destroy the discriminative visual cues residing in the subtle regions. In this paper, we propose diversifying the training data at the feature-level to alleviate the discriminative region loss problem. Specifically, we produce diversified augmented samples by translating image features along semantically meaningful directions. The semantic directions are estimated with a covariance prediction network, which predicts a sample-wise covariance matrix to adapt to the large intra-class variation inherent in fine-grained images. Furthermore, the covariance prediction network is jointly optimized with the classification network in a meta-learning manner to alleviate the degenerate solution problem. Experiments on four competitive fine-grained recognition benchmarks (CUB-200-2011, Stanford Cars, FGVC Aircrafts, NABirds) demonstrate that our method significantly improves the generalization performance on several popular classification networks (e.g., ResNets, DenseNets, EfficientNets, RegNets and ViT). Combined with a recently proposed method, our semantic data augmentation approach achieves state-of-the-art performance on the CUB-200-2011 dataset. Source code is available at https://github.com/LeapLabTHU/LearnableISDA.

The development of vision models for real-world applications is hindered by the challenge of annotated data scarcity, which has necessitated the adoption of dataefficient visual learning techniques such as semi-supervised learning. Unfortunately, the prevalent cross-entropy supervision is limited by its focus on category discrimination while disregarding the semantic connection between concepts, which ultimately results in the suboptimal exploitation of scarce labeled data. To address this issue, this paper presents a novel approach that seeks to leverage linguistic knowledge for data-efficient visual learning. The proposed approach, BorLan, Borrows knowledge from off-theshelf pretrained Language models that are already endowed with rich semantics extracted from large corpora, to compensate the semantic deficiency due to limited annotation in visual training. Specifically, we design a distribution alignment objective, which guides the vision model to learn both semantic-aware and domain-agnostic representations for the task through linguistic knowledge. One significant advantage of this paradigm is its flexibility in combining various visual and linguistic models. Extensive experiments on semi-supervised learning, single domain generalization and few-shot learning validate its effectiveness. Code is available at https://github.com/BIT-DA/BorLan.

Early exiting has become a promising approach to improving the inference efficiency of deep networks. By structuring models with multiple classifiers (exits), predictions for "easy" samples can be generated at earlier exits, negating the need for executing deeper layers. Current multi-exit networks typically implement linear classifiers at intermediate layers, compelling low-level features to encapsulate high-level semantics. This sub-optimal design invariably undermines the performance of later exits. In this paper, we propose Dynamic Perceiver (Dyn-Perceiver) to decouple the feature extraction procedure and the early classification task with a novel dual-branch architecture. A feature branch serves to extract image features, while a classification branch processes a latent code assigned for classification tasks. Bi-directional cross-attention layers are established to progressively fuse the information of both branches. Early exits are placed exclusively within the classification branch, thus eliminating the need for linear separability in low-level features. Dyn-Perceiver constitutes a versatile and adaptable framework that can be built upon various architectures. Experiments on image classification, action recognition, and object detection demonstrate that our method significantly improves the inference efficiency of different backbones, outperforming numerous competitive approaches across a broad range of computational budgets. Evaluation on both CPU and GPU platforms substantiate the superior practical efficiency of Dyn-Perceiver. Code is available at https://www.github.com/LeapLabTHU/Dynamic_Perceiver.

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.

Rotated object detection aims to identify and locate objects in images with arbitrary orientation. In this scenario, the oriented directions of objects vary considerably across different images, while multiple orientations of objects exist within an image. This intrinsic characteristic makes it challenging for standard backbone networks to extract high-quality features of these arbitrarily orientated objects. In this paper, we present Adaptive Rotated Convolution (ARC) module to handle the afore-mentioned challenges. In our ARC module, the convolution kernels rotate adaptively to extract object features with varying orientations in different images, and an efficient conditional computation mechanism is introduced to accommodate the large orientation variations of objects within an image. The two designs work seamlessly in rotated object detection problem. Moreover, ARC can conveniently serve as a plug-and-play module in various vision backbones to boost their representation ability to detect oriented objects accurately. Experiments on commonly used benchmarks (DOTA and HRSC2016) demonstrate that equipped with our proposed ARC module in the backbone network, the performance of multiple popular oriented object detectors is significantly improved (e.g. +3.03% mAP on Rotated RetinaNet and +4.16% on CFA). Combined with the highly competitive method Oriented R-CNN, the proposed approach achieves state-of-the-art performance on the DOTA dataset with 81.77% mAP. Code is available at https://github.com/LeapLabTHU/ARC.

The superior performance of modern deep networks usually comes with a costly training procedure. This paper presents a new curriculum learning approach for the efficient training of visual backbones (e.g., vision Transformers). Our work is inspired by the inherent learning dynamics of deep networks: we experimentally show that at an earlier training stage, the model mainly learns to recognize some ‘easier-to-learn’ discriminative patterns within each example, e.g., the lower-frequency components of images and the original information before data augmentation. Driven by this phenomenon, we propose a curriculum where the model always leverages all the training data at each epoch, while the curriculum starts with only exposing the ‘easier-to-learn’ patterns of each example, and introduces gradually more difficult patterns. To implement this idea, we 1) introduce a cropping operation in the Fourier spectrum of the inputs, which enables the model to learn from only the lower-frequency components efficiently, 2) demonstrate that exposing the features of original images amounts to adopting weaker data augmentation, and 3) integrate 1) and 2) and design a curriculum learning schedule with a greedy-search algorithm. The resulting approach, EfficientTrain, is simple, general, yet surprisingly effective. As an off-the-shelf method, it reduces the wall-time training cost of a wide variety of popular models (e.g., ResNet, ConvNeXt, DeiT, PVT, Swin, and CSWin) by > 1.5× on ImageNet-1K/22K without sacrificing accuracy. It is also effective for self-supervised learning (e.g., MAE). Code is available at https://github.com/LeapLabTHU/EfficientTrain.

Spatial redundancy widely exists in visual recognition tasks, i.e., discriminative features in an image or video frame usually correspond to only a subset of pixels, while the remaining regions are irrelevant to the task at hand. Therefore, static models which process all the pixels with an equal amount of computation result in considerable redundancy in terms of time and space consumption. In this paper, we formulate the image recognition problem as a sequential coarse-to-fine feature learning process, mimicking the human visual system. Specifically, the proposed Glance and Focus Network (GFNet) first extracts a quick global representation of the input image at a low resolution scale, and then strategically attends to a series of salient (small) regions to learn finer features. The sequential process naturally facilitates adaptive inference at test time, as it can be terminated once the model is sufficiently confident about its prediction, avoiding further redundant computation. It is worth noting that the problem of locating discriminant regions in our model is formulated as a reinforcement learning task, thus requiring no additional manual annotations other than classification labels. GFNet is general and flexible as it is compatible with any off-the-shelf backbone models (such as MobileNets, EfficientNets and TSM), which can be conveniently deployed as the feature extractor. Extensive experiments on a variety of image classification and video recognition tasks and with various backbone models demonstrate the remarkable efficiency of our method. For example, it reduces the average latency of the highly efficient MobileNet-V3 on an iPhone XS Max by 1.3x without sacrificing accuracy. Code and pre-trained models are available at https://github.com/blackfeather-wang/GFNet-Pytorch.

Recent works have shown that the computational efficiency of video recognition can be significantly improved by reducing the spatial redundancy. As a representative work, the adaptive focus method (AdaFocus) has achieved a favorable trade-off between accuracy and inference speed by dynamically identifying and attending to the informative regions in each video frame. However, AdaFocus requires a complicated three-stage training pipeline (involving reinforcement learning), leading to slow convergence and is unfriendly to practitioners. This work reformulates the training of AdaFocus as a simple one-stage algorithm by introducing a differentiable interpolation-based patch selection operation, enabling efficient end-to-end optimization. We further present an improved training scheme to address the issues introduced by the one-stage formulation, including the lack of supervision, input diversity and training stability. Moreover, a conditional-exit technique is proposed to perform temporal adaptive computation on top of AdaFocus without additional training. Extensive experiments on six benchmark datasets (i.e., ActivityNet, FCVID, Mini-Kinetics, Something-Something V1&V2, and Jester) demonstrate that our model significantly outperforms the original AdaFocus and other competitive baselines, while being considerably more simple and efficient to train. Code is available at https://github.com/LeapLabTHU/AdaFocusV2.

Extensive studies on Unsupervised Domain Adaptation (UDA) have propelled the deployment of deep learning from limited experimental datasets into real-world unconstrained domains. Most UDA approaches align features within a common embedding space and apply a shared classifier for target prediction. However, since a perfectly aligned feature space may not exist when the domain discrepancy is large, these methods suffer from two limitations. First, the coercive domain alignment deteriorates target domain discriminability due to lacking target label supervision. Second, the source-supervised classifier is inevitably biased to source data, thus it may underperform in target domain. To alleviate these issues, we propose to simultaneously conduct feature alignment in two individual spaces focusing on different domains, and create for each space a domain-oriented classifier tailored specifically for that domain. Specifically, we design a Domain-Oriented Transformer (DOT) that has two individual classification tokens to learn different domain-oriented representations, and two classifiers to preserve domain-wise discriminability. Theoretical guaranteed contrastive-based alignment and the source-guided pseudo-label refinement strategy are utilized to explore both domain-invariant and specific information. Comprehensive experiments validate that our method achieves state-of-the-art on several benchmarks. Code is released at https://github.com/BIT-DA/Domain-Oriented-Transformer.

Dynamic neural network is an emerging research topic in deep learning. Compared to static models which have fixed computational graphs and parameters at the inference stage, dynamic networks can adapt their structures or parameters to different inputs, leading to notable advantages in terms of accuracy, computational efficiency, adaptiveness, etc. In this survey, we comprehensively review this rapidly developing area by dividing dynamic networks into three main categories: 1) sample-wise dynamic models that process each sample with data-dependent architectures or parameters; 2) spatial-wise dynamic networks that conduct adaptive computation with respect to different spatial locations of image data; and 3) temporal-wise dynamic models that perform adaptive inference along the temporal dimension for sequential data such as videos and texts. The important research problems of dynamic networks, e.g., architecture design, decision making scheme, optimization technique and applications, are reviewed systematically. Finally, we discuss the open problems in this field together with interesting future research directions.

Data augmentation is widely known as a simple yet surprisingly effective technique for regularizing deep networks. Conventional data augmentation schemes, e.g., flipping, translation or rotation, are low-level, data-independent and class-agnostic operations, leading to limited diversity for augmented samples. To this end, we propose a novel semantic data augmentation algorithm to complement traditional approaches. The proposed method is inspired by the intriguing property that deep networks are effective in learning linearized features, i.e., certain directions in the deep feature space correspond to meaningful semantic transformations, e.g., changing the background or view angle of an object. Based on this observation, translating training samples along many such directions in the feature space can effectively augment the dataset for more diversity. To implement this idea, we first introduce a sampling based method to obtain semantically meaningful directions efficiently. Then, an upper bound of the expected cross-entropy (CE) loss on the augmented training set is derived by assuming the number of augmented samples goes to infinity, yielding a highly efficient algorithm. In fact, we show that the proposed implicit semantic data augmentation (ISDA) algorithm amounts to minimizing a novel robust CE loss, which adds minimal extra computational cost to a normal training procedure. In addition to supervised learning, ISDA can be applied to semi-supervised learning tasks under the consistency regularization framework, where ISDA amounts to minimizing the upper bound of the expected KL-divergence between the augmented features and the original features. Although being simple, ISDA consistently improves the generalization performance of popular deep models (e.g., ResNets and DenseNets) on a variety of datasets, i.e., CIFAR-10, CIFAR-100, SVHN, ImageNet, and Cityscapes. Code for reproducing our results is available at https://github.com/blackfeather-wang/ISDA-for-Deep-Networks.

Real-world training data usually exhibits long-tailed distribution, where several majority classes have a significantly larger number of samples than the remaining minority classes. This imbalance degrades the performance of typical supervised learning algorithms designed for balanced training sets. In this paper, we address this issue by augmenting minority classes with a recently proposed implicit semantic data augmentation (ISDA) algorithm [37], which produces diversified augmented samples by translating deep features along many semantically meaningful directions. Importantly, given that ISDA estimates the class-conditional statistics to obtain semantic directions, we find it ineffective to do this on minority classes due to the insufficient training data. To this end, we propose a novel approach to learn transformed semantic directions with meta-learning automatically. In specific, the augmentation strategy during training is dynamically optimized, aiming to minimize the loss on a small balanced validation set, which is approximated via a meta update step. Extensive empirical results on CIFAR-LT-10/100, ImageNet-LT, and iNaturalist 2017/2018 validate the effectiveness of our method.

Domain adaptation has been widely explored by transferring the knowledge from a label-rich source domain to a related but unlabeled target domain. Most existing domain adaptation algorithms attend to adapting feature representations across two domains with the guidance of a shared source-supervised classifier. However, such classifier limits the generalization ability towards unlabeled target recognition. To remedy this, we propose a Transferable Semantic Augmentation (TSA) approach to enhance the classifier adaptation ability through implicitly generating source features towards target semantics. Specifically, TSA is inspired by the fact that deep feature transformation towards a certain direction can be represented as meaningful semantic altering in the original input space. Thus, source features can be augmented to effectively equip with target semantics to train a more transferable classifier. To achieve this, for each class, we first use the inter-domain feature mean difference and target intra-class feature covariance to construct a multivariate normal distribution. Then we augment source features with random directions sampled from the distribution class-wisely. Interestingly, such source augmentation is implicitly implemented through an expected transferable cross-entropy loss over the augmented source distribution, where an upper bound of the expected loss is derived and minimized, introducing negligible computational overhead. As a light-weight and general technique, TSA can be easily plugged into various domain adaptation methods, bringing remarkable improvements. Comprehensive experiments on cross-domain benchmarks validate the efficacy of TSA.

Reusing features in deep networks through dense connectivity is an effective way to achieve high computational efficiency. The recent proposed CondenseNet [14] has shown that this mechanism can be further improved if redundant features are removed. In this paper, we propose an alternative approach named sparse feature reactivation (SFR), aiming at actively increasing the utility of features for reusing. In the proposed network, named CondenseNetV2, each layer can simultaneously learn to 1) selectively reuse a set of most important features from preceding layers; and 2) actively update a set of preceding features to increase their utility for later layers. Our experiments show that the proposed models achieve promising performance on image classification (ImageNet and CIFAR) and object detection (MS COCO) in terms of both theoretical efficiency and practical speed.

In this paper, we explore the spatial redundancy in video recognition with the aim to improve the computational efficiency. It is observed that the most informative region in each frame of a video is usually a small image patch, which shifts smoothly across frames. Therefore, we model the patch localization problem as a sequential decision task, and propose a reinforcement learning based approach for efficient spatially adaptive video recognition (AdaFocus). In specific, a light-weighted ConvNet is first adopted to quickly process the full video sequence, whose features are used by a recurrent policy network to localize the most task-relevant regions. Then the selected patches are inferred by a high-capacity network for the final prediction. During offline inference, once the informative patch sequence has been generated, the bulk of computation can be done in parallel, and is efficient on modern GPU devices. In addition, we demonstrate that the proposed method can be easily extended by further considering the temporal redundancy, e.g., dynamically skipping less valuable frames. Extensive experiments on five benchmark datasets, i.e., ActivityNet, FCVID, MiniKinetics, Something-Something V1&V2, demonstrate that our method is significantly more efficient than the competitive baselines. Code is available at https://github.com/blackfeather-wang/AdaFocus.

Vision Transformers (ViT) have achieved remarkable success in large-scale image recognition. They split every 2D image into a fixed number of patches, each of which is treated as a token. Generally, representing an image with more tokens would lead to higher prediction accuracy, while it also results in drastically increased computational cost. To achieve a decent trade-off between accuracy and speed, the number of tokens is empirically set to 16x16 or 14x14. In this paper, we argue that every image has its own characteristics, and ideally the token number should be conditioned on each individual input. In fact, we have observed that there exist a considerable number of “easy” images which can be accurately predicted with a mere number of 4x4 tokens, while only a small fraction of “hard” ones need a finer representation. Inspired by this phenomenon, we propose a Dynamic Transformer to automatically configure a proper number of tokens for each input image. This is achieved by cascading multiple Transformers with increasing numbers of tokens, which are sequentially activated in an adaptive fashion at test time, i.e., the inference is terminated once a sufficiently confident prediction is produced. We further design efficient feature reuse and relationship reuse mechanisms across different components of the Dynamic Transformer to reduce redundant computations. Extensive empirical results on ImageNet, CIFAR-10, and CIFAR-100 demonstrate that our method significantly outperforms the competitive baselines in terms of both theoretical computational efficiency and practical inference speed. Code and pre-trained models (based on PyTorch and MindSpore) are available at https://github.com/blackfeather-wang/Dynamic-Vision-Transformer and https://github.com/blackfeather-wang/Dynamic-Vision-Transformer-MindSpore.

The accuracy of deep convolutional neural networks (CNNs) generally improves when fueled with high resolution images. However, this often comes at a high computational cost and high memory footprint. Inspired by the fact that not all regions in an image are task-relevant, we propose a novel framework that performs efficient image classification by processing a sequence of relatively small inputs, which are strategically selected from the original image with reinforcement learning. Such a dynamic decision process naturally facilitates adaptive inference at test time, i.e., it can be terminated once the model is sufficiently confident about its prediction and thus avoids further redundant computation. Notably, our framework is general and flexible as it is compatible with most of the state-of-the-art light-weighted CNNs (such as MobileNets, EfficientNets and RegNets), which can be conveniently deployed as the backbone feature extractor. Experiments on ImageNet show that our method consistently improves the computational efficiency of a wide variety of deep models. For example, it further reduces the average latency of the highly efficient MobileNet-V3 on an iPhone XS Max by 20% without sacrificing accuracy. Code and pre-trained models are available at https://github.com/blackfeather-wang/GFNet-Pytorch.

In this paper, we propose a novel implicit semantic data augmentation (ISDA) approach to complement traditional augmentation techniques like flipping, translation or rotation. Our work is motivated by the intriguing property that deep networks are surprisingly good at linearizing features, such that certain directions in the deep feature space correspond to meaningful semantic transformations, e.g., adding sunglasses or changing backgrounds. As a consequence, translating training samples along many semantic directions in the feature space can effectively augment the dataset to improve generalization. To implement this idea effectively and efficiently, we first perform an online estimate of the covariance matrix of deep features for each class, which captures the intra-class semantic variations. Then random vectors are drawn from a zero-mean normal distribution with the estimated covariance to augment the training data in that class. Importantly, instead of augmenting the samples explicitly, we can directly minimize an upper bound of the expected cross-entropy (CE) loss on the augmented training set, leading to a highly efficient algorithm. In fact, we show that the proposed ISDA amounts to minimizing a novel robust CE loss, which adds negligible extra computational cost to a normal training procedure. Although being simple, ISDA consistently improves the generalization performance of popular deep models (ResNets and DenseNets) on a variety of datasets, e.g., CIFAR-10, CIFAR-100 and ImageNet. Code for reproducing our results are available at https://github.com/blackfeather-wang/ISDA-for-Deep-Networks.