论文检索

Chuang Liu, Yichao Cao, Xiu Su, Haogang Zhu

Test-time adaptation (TTA) has emerged as a promising solution to address real world domain shifts in medical image segmentation. Current approaches adapt by updating or regularizing a pre-trained source model. However, they face two major issues: (i) the source models on which they rely are prone to overfitting under domain shifts; (ii) in dynamic continual testing scenarios, error accumulation and class forgetting are further exacerbated. To overcome these limitations, we propose TanGo, a novel framework that combines Training to adapt with Foundation Guidance and continual style calibration. During training, TanGo learns generalization priors from vision foundation models (VFMs) through distribution-wise consistency learning. We incorporate stable low-frequency representations from a frozen encoder of VFMs as priors to guide the source model, constraining its output feature distribution to yield a more generalizable feature space. At test time, we propose an instance-wise style calibration method that uses a learnable data decorator to shift dynamic test images back toward an enhanced source-like distribution. Subsequently, a series of constraints is applied to draw the decorated test samples closer to the enhanced source distribution while maintaining their semantic integrity, thus boosting the continual adaptability of the source model. Extensive experiments on multiple medical image segmentation tasks demonstrate that TanGo achieves state-of-the-art performance.

Jianing Qian, Qinhe Peng, Emmanuel Panov, Leonor Fermoselle, Dinesh Jayaraman, Bernadette Bucher, Tarik Kelestemur

Imitation learning enables robots to learn how to execute tasks via observation. However, real-world environments like homes and offices are often severely partially observed due to their large spatial scales. In addition, many tasks involve executing a series of subtasks requiring autonomous robots to reason over extended time horizons. To address these challenges, we propose using scene graphs as an explicit and structured memory mechanism in imitation learning. By maintaining a dynamic scene graph that captures object-centric relationships and their evolution over time, our method allows the agent to retain relevant historical context during task execution to efficiently reason over incrementally accrued scene information. Our experiments on simulated mobile manipulation and real-world tabletop manipulation demonstrate that our approach substantially improves policy performance, particularly in settings that demand long-term reasoning and robust generalization under partial observability.

Idan Yankelev, Edita Grolman, Yarin Yerushalmi Levi, Amit Giloni, Omer Hofman, Toshiya Shimizu, Yuval Elovici, Asaf Shabtai

Adversarial patch attacks pose a significant threat to the reliability of object detection (OD) models, particularly in real-time security applications. Although several defenses have been proposed, they often suffer from two limitations: 1) reduced performance on benign images, and 2) impractical processing time for real-time OD applications. In this paper, we present AntiStyler, a novel and rapid defense against adversarial patches. Given an input image, AntiStyler identifies and masks pixels that exhibit a "random" style associated with adversarial attacks and uses a series of spatial filters to enhance the mask and remove unwanted noise, efficiently masking adversarial patches. AntiStyler features model-, patch-, and attack-agnostic capabilities and does not require any training, making it a fully agnostic zero-shot defense against adversarial patch attacks. Our evaluation on the COCO, INRIA, Superstore, and APRICOT datasets, with both digital and physical attacks, demonstrates AntiStyler's state-of-the-art robustness (improving adversarial performance by 8-15 mAP%) without compromising the original performance on benign images. Additionally, unlike most existing defenses, AntiStyler can process 10-12 frames per second (FPS), making it efficient and relevant for real-time OD applications.

Dongchen Han, Yining Li, Tianyu Li, Zixuan Cao, Ziming Wang, Jun Song, Yu Cheng, Bo Zheng, Gao Huang

Test-Time Training (TTT) has recently emerged as a promising direction for efficient sequence modeling. TTT reformulates attention operation as an online learning problem, constructing a compact inner model from key-value pairs at test time. This reformulation opens a rich and flexible design space while achieving linear computational complexity. However, crafting a powerful visual TTT design remains challenging: fundamental choices for the inner module and inner training lack comprehensive understanding and practical guidelines. To bridge this critical gap, in this paper, we present a systematic empirical study of TTT designs for visual sequence modeling. From a series of experiments and analyses, we distill six practical insights that establish design principles for effective visual TTT and illuminate paths for future improvement. These findings culminate in the Vision Test-Time Training (ViT^3) model, a pure TTT architecture that achieves linear complexity and parallelizable computation. We evaluate ViT^3 across diverse visual tasks, including image classification, image generation, object detection, and semantic segmentation. Results show that ViT^3 consistently matches or outperforms advanced linear-complexity models (e.g., Mamba and linear attention variants) and effectively narrows the gap to highly optimized vision Transformers. We hope this study and the ViT^3 baseline can facilitate future work on visual TTT models. Code: github.com/LeapLabTHU/ViTTT.

Zhengpeng Feng, Clement Atzberger, Sadiq Jaffer, Jovana Knezevic, Silja Sormunen, Robin Young, Madeline C. Lisaius, Markus Immitzer, Toby Jackson, James Ball 等

Satellite Earth-observation (EO) time series in the optical and microwave ranges are often irregular due to orbital patterns and cloud obstruction, and while compositing addresses these issues, it loses critical phenological information. To overcome this, we present TESSERA, a pixel-wise foundation model for multi-modal (Sentinel-1/2) EO time series that learns robust, label-efficient embeddings. During training, TESSERA uses Barlow Twins and sparse random temporal sampling to enforce invariance to the selection of valid observations, aided by two key regularizers: global shuffling to decorrelate spatial neighborhoods and mix-based regulation for invariance under extreme sparsity. We find that for diverse classification, segmentation, and regression tasks, TESSERA embeddings deliver state-of-the-art accuracy with high label efficiency, often requiring only a small task head and minimal computation. To democratize access, adhere to FAIR principles, and simplify use, we release global, annual, 10m, pixel-wise int8 embeddings together with open weights/code and lightweight adaptation heads, providing practical tooling for large-scale retrieval and inference at planetary scale. All code and data are available at https://github.com/ucam-eo/tessera.

Pablo Ruiz-Ponce, Sergio Escalera, José García-Rodríguez, Jiankang Deng, Rolandos Alexandros Potamias

Generating realistic human-human interactions is a challenging task that requires not only high-quality individual body and hand motions, but also coherent coordination among all interactants. Due to limitations in available data and increased learning complexity, previous methods tend to ignore hand motions, limiting the realism and expressivity of the interactions. Additionally, current diffusion-based approaches generate entire motion sequences simultaneously, limiting their ability to capture the reactive and adaptive nature of human interactions. To address these limitations, we introduce Interact2Ar, the first end-to-end text-conditioned autoregressive diffusion model for generating full-body, human-human interactions. Interact2Ar incorporates detailed hand kinematics through dedicated parallel branches, enabling high-fidelity full-body generation. Furthermore, we introduce an autoregressive pipeline coupled with a novel memory technique that facilitates adaptation to the inherent variability of human interactions using efficient large context windows. The adaptability of our model enables a series of downstream applications, including temporal motion composition, real-time adaptation to disturbances, and extension beyond dyadic to multi-person scenarios. To validate the generated motions, we introduce a set of robust evaluators and extended metrics designed specifically for assessing full-body interactions. Through quantitative and qualitative experiments, we demonstrate the state-of-the-art performance of Interact2Ar.

Yushi Huang, Zining Wang, Zhihang Yuan, Yifu Ding, Ruihao Gong, Jinyang Guo, Xianglong Liu, Jun Zhang

Mixture-of-Experts (MoE) Multimodal large language models (MLLMs) excel at vision-language tasks, but they suffer from high computational inefficiency. To reduce inference overhead, expert skipping methods have been proposed to deactivate redundant experts based on the current input tokens. However, we find that applying these methods--originally designed for unimodal large language models (LLMs)--to MLLMs results in considerable performance degradation. This is primarily because such methods fail to account for the heterogeneous contributions of experts across MoE layers and modality-specific behaviors of tokens within these layers. Motivated by these findings, we propose MoDES, the first training-free framework that adaptively skips experts to enable efficient and accurate MoE MLLM inference. It incorporates a globally-modulated local gating (GMLG) mechanism that integrates global layer-wise importance into local routing probabilities to accurately estimate per-token expert importance. A dual-modality thresholding (DMT) method is then applied, which processes tokens from each modality separately, to derive the skipping schedule. To set the optimal thresholds, we introduce a frontier search algorithm that exploits monotonicity properties, cutting convergence time from several days to a few hours. Extensive experiments for 3 model series across 13 benchmarks demonstrate that MoDES far outperforms previous approaches. For instance, when skipping 88% experts for Qwen3-VL-MoE-30B-A3B-Instruct, the performance boost is up to **10.67%** (97.33% vs. 86.66%). Furthermore, MoDES significantly enhances inference speed, improving the prefilling time by **2.16x** and the decoding time by **1.26x**.

Zixun Wang

The F-measure is a widely used metric in multi-label classification, where multiple labels are predicted simultaneously for a single instance. The optimal prediction rule for F-measure requires estimating q^2+1 probabilities, where q is the number of labels. Existing approaches train q multinomial estimators (multi-class classifiers) to directly estimate these probabilities, followed by a matrix multiplication for making predictions. However, this method has two major drawbacks. First, the matrix multiplication incurs a time complexity of \mathcal O (q^3), which becomes computationally expensive for large q. Second, training multinomial estimators is challenging due to the sparsity of the underlying distributions, which results from the inherent imbalance in multi-label datasets and is further exacerbated by the label transformation required by the method itself. In this paper, we first demonstrate that matrix multiplication can be reformulated as a series of convolutions by exploiting a special structure in the matrix. These convolutions can then be efficiently computed using the Fast Fourier Transform (FFT), reducing the time complexity to \mathcal O (q^2\log q). To avoid multinomial label transformation, we propose an indirect sampling-then-estimation approach to estimate the required probabilities. This method trains only q binary estimators instead of multinomial ones, thereby alleviating the sparsity issue, simplifying the training process, and improving performance. We provide theoretical guarantees for the consistency of the proposed sampling-based method and demonstrate its effectiveness through extensive experiments on diverse datasets. The code is available in: https://github.com/ZixunWang/MLC-F1-Sampling.

Yu Gao, Lutong Su, Ruixiang Huang, Tianji Jiang, Jiadong Tang, Yufeng Yue, Yi Yang

High-quality 3D scene representation in radiance fields relies on accurate camera poses which are often difficult to acquire in real-world scenarios. An effective solution is to use RGB images for the joint optimization of radiance fields and camera poses, an approach that has been well explored in NeRF series methods. However, unlike NeRF, joint optimization in 3D Gaussian Splatting (3DGS) often requires additional regularization or prior spatial knowledge to reach comparable performance. To eliminate these dependencies, we introduce Energy-GS, a pose-aware Gaussian splatting framework that jointly optimizes scene representation and camera poses using only RGB images. We observe that pose gradients in joint optimization are unstable due to the point-based rendering mechanism. Furthermore, unlike NeRF's spatial sampling framework that enables coarse-to-fine pose alignment, rasterization-based 3DGS lacks controllable sampling and thus cannot support progressive pose refinement. To address these challenges, we redesign the optimization strategy of Gaussian primitives and introduce an image-energy-guided constraint that encourages progressive alignment of camera poses. Experiments on both synthetic and real-world datasets show that Energy-GS can effectively optimize the scene reconstruction and resolve camera pose misalignment at the same time. Benefiting from reliance on only RGB images, we believe this work provides promising insights for visual localization and dense mapping applications such as SLAM. The Code is available at \href https://github.com/SkylerGao/ENGS https://github.com/SkylerGao/ENGS

Shasha Han, Chong Li, Xinning Wang, Xuebo Li

YOLO series lead object detection with superior accuracy and speed. However, both convolutional and self-attention based architectures suffer from parameter redundancy and insufficient computational efficiency. Existing lightweight methods excessively pursue speed while ignoring the loss of important information during feature extraction and spatial transformation across different stages. Thus, effective lightweighting is crucial for detection performance. We propose YOLO-ULM, an ultra-lightweight real-time detector that achieves accelerated inference while preserving high accuracy. We innovatively design a variety of dual efficiency- and accuracy-driven modules, including efficient feature aggregation and parallel downsampling modules, as well as a more focused complete-IoU loss function. To validate our approach, we train it from scratch on COCO dataset without pretrained weights. By refining backbone parameters, we extend it to YOLO-ULM-Turbo for accelerated inference. YOLO-ULM surpasses state-of-the-art real-time detectors like YOLOv11/YOLOv12/YOLOv13 and RT-DETR. On a T4 GPU, YOLO-ULM-N achieves 41.6% mAP with an inference latency of 1.52 ms, outperforming YOLOv11-N (2.2%\uparrow) and YOLOv12-N (1.0%\uparrow). YOLO-ULM-S exceeds RT-DETR-R18 by 1.6% mAP with 64.7% fewer FLOPs and 63% fewer parameters. YOLO-ULM-L / X surpass YOLOv13-L / X by 0.7% and 0.8% respectively in mAP. YOLO-ULM-Turbo matches YOLOv12-Turbo's performance but uses less computation, with Turbo-N variant achieving 0.3% higher mAP and 16% fewer parameters than YOLOv12-Turbo-N.

Hanjie Xu, Yuanxing Duan, Qiyu Dai, Ge Li, Baoquan Chen, He Wang

We address the challenge of reconstructing long dynamic scenes from multi-view videos in a storage-efficient manner. Recent advances in Gaussian Splatting and its extensions to dynamic scenes have demonstrated impressive visual quality, but remain limited to short duration (<10 s), large storage size (>500 MB), and high GPU VRAM usage.To overcome these limitations, we introduce Layered 4D-Rotor Gaussian Splatting (L4DRotorGS), a novel compressed representation designed for long dynamic scenes. Our approach integrates a layered 4D representation, efficient training, and effective compression into a unified framework. Specifically, 4D Gaussians are first organized into layers based on their temporal extents and then partitioned into discrete temporal buckets. This structure allows for selective access and rendering of only the necessary subsets of 4D Gaussians, substantially reducing GPU memory requirements.To further compress the representation, we apply a series of techniques, Factorized Covariance Quantization, Layered Compression, and Residual Codebook Quantization, achieving a compression ratio of up to 22.3x while preserving high visual fidelity.We implement a highly optimized C++/CUDA framework for efficient training, compression, and real-time rendering, achieving over 500 FPS on an RTX 3090 GPU. Extensive experiments demonstrate the superior storage efficiency, visual quality, and rendering speed of L4DRotorGS, consistently outperforming prior methods in both quantitative metrics and perceptual quality on real-world long dynamic scenes.

Jun Zhang, Teng Wang, Yuying Ge, Yixiao Ge, Xinhao Li, Limin Wang

This paper does not introduce a novel method but instead establishes a straightforward, incremental, yet essential baseline for video temporal grounding (VTG), a core capability in video understanding. While multimodal large language models (MLLMs) excel at various video understanding tasks, the recipes for optimizing them for VTG remain under-explored. In this paper, we present TimeLens, a systematic investigation into building MLLMs with strong VTG ability, along two primary dimensions: data quality and algorithmic design. We first expose critical quality issues in existing VTG benchmarks and introduce TimeLens-Bench, comprising meticulously re-annotated versions of three popular benchmarks with strict quality criteria. Our analysis reveals dramatic model re-rankings compared to legacy benchmarks, confirming the unreliability of prior evaluation standards. We also address noisy training data through an automated re-annotation pipeline, yielding TimeLens-100K, a large-scale, high-quality training dataset. Building on our data foundation, we conduct in-depth explorations of algorithmic design principles, yielding a series of meaningful insights and effective yet efficient practices. These include interleaved textual encoding for time representation, a thinking-free reinforcement learning with verifiable rewards (RLVR) approach as the training paradigm, and carefully designed recipes for RLVR training. These efforts culminate in TimeLens models, a family of MLLMs with state-of-the-art VTG performance among open-source models and even surpass proprietary models such as GPT-5 and Gemini-2.5-Flash. All codes, data, and models will be released to facilitate future research.

Nicolas Houdré, Diego Marcos, Hugo Riffaud de Turckheim, Dino Ienco, Laurent Wendling, Camille Kurtz, Sylvain Lobry

Earth observation (EO) data spans a wide range of spatial, spectral, and temporal resolutions, from high-resolution optical imagery to low resolution multispectral products or radar time series. While recent foundation models have improved multimodal integration for learning meaningful representations, they often expect fixed input resolutions or are based on sensor-specific encoders limiting generalization across heterogeneous EO modalities. To overcome these limitations we introduce RAMEN, a resolution-adjustable multimodal encoder that learns a shared visual representation across EO data in a fully sensor-agnostic manner. RAMEN treats the modality and spatial and temporal resolutions as key input data features, enabling coherent analysis across modalities within a unified latent space. Its main methodological contribution is to define spatial resolution as a controllable output parameter, giving users direct control over the desired level of detail at inference and allowing explicit trade-offs between spatial precision and computational cost. We train a single, unified transformer encoder reconstructing masked multimodal EO data drawn from diverse sources, ensuring generalization across sensors and resolutions. Once pretrained, RAMEN transfers effectively to both known and unseen sensor configurations and outperforms larger state-of-the-art models on the community-standard PANGAEA benchmark, containing various multi-sensor and multi-resolution downstream tasks. Our code and pretrained model will be released upon acceptance.

Lei Zhou, Haoyu Wu, Akshat Dave, Dimitris Samaras

We introduce a test-time framework for multiview Transformers (MVTs) that incorporates priors (e.g., camera poses, intrinsics, and depth) to improve 3D tasks without retraining or modifying pre-trained image-only networks. Rather than feeding priors into the architecture, we cast them as constraints on the predictions and optimize the network at inference time. The optimization loss consists of a self-supervised objective and prior penalty terms. The self-supervised objective captures the compatibility among multi-view predictions and is implemented using photometric or geometric loss between renderings from other views and each view itself. Any available priors are converted into penalty terms on the corresponding output modalities. Across a series of 3D vision benchmarks, including point map estimation and camera pose estimation, our method consistently improves performance over base MVTs by a large margin. On the ETH3D, 7-Scenes, and NRGBD datasets, our method reduces the point-map distance error by more than half compared with the base image-only models. Our method also outperforms retrained prior-aware feed-forward methods, demonstrating the effectiveness of our test-time constrained optimization (TCO) framework for incorporating priors into 3D vision tasks. The code is available at https://github.com/cvlab-stonybrook/TCO.

Yiyi Cai, Yuhan Wu, Kunhang Li, You Zhou, Bo Zheng, Haiyang Liu

We present FloodDiffusion, a new framework for text-driven, streaming human motion generation. Given time-varying text prompts, FloodDiffusion generates text-aligned, seamless motion sequences with real-time latency.Unlike existing methods that rely on chunk-by-chunk or auto-regressive model with diffusion head, we adopt a diffusion forcing framework to model this time-series generation task under time-varying control events.We find that a straightforward implementation of vanilla diffusion forcing (as proposed for video models) fails to model real motion distributions. We demonstrate that to guarantee modeling the output distribution, the vanilla diffusion forcing must be tailored to: (i) train with a bi-directional attention instead of casual attention; (ii) implement a lower triangular time scheduler instead of a random one; (iii) utilize a continues time-varying way to introduce text conditioning.With these improvements, we demonstrate in the first time that the diffusion forcing-based framework achieves state-of-the-art performance on the streaming motion generation task, reaching an FID of 0.057 on the HumanML3D benchmark. Models, code, and weights are available.

Bishal Ranjan Swain, Kyung Joo Cheoi, Jaepil Ko

Continuous-time neural networks provide adaptive dynamics, but rely on a single hidden state to encode both fast input fluctuations and longer-term context. This shared representation forces rapidly changing inputs to overwrite slower contextual signals, causing the model to lose past information as new observations arrive. In contrast, biological perceptual systems maintain stable behaviour under evolving sensory input by integrating ongoing signals with stored associative patterns rather than relying on a single evolving state.Motivated by this distinction, we study a simple coupling of Liquid Time-Constant Networks (LTCs) with a Modern Hopfield Network (MHN) that serves as a content-addressable memory. At each time step, the liquid state is projected into a query, the MHN retrieves a memory vector, and the two representations are concatenated before a readout layer. We analyse this coupling under standard norm and Lipschitz assumptions and show that the combined representation remains bounded. We further show that the retrieval map contracts gradients for parameters upstream of the memory query, which provides a mechanism for reducing curvature in the loss landscape.On public time-series benchmarks, the coupled LTC-MHN model improves mean accuracy by 2.3% over competitive recurrent and continuous-time baselines and reduces the estimated Hessian trace by about an order of magnitude relative to a standalone LTC encoder, with the largest gains on classification tasks and competitive performance on a regression task. Qualitative analyses of training curves, loss landscapes, and latent embeddings support the interpretation that Hopfield retrieval smooths optimization and encourages more compact, linearly separable class manifolds. Code will be released upon publication.

Khushboo Mishra, Varun Trivedi, Tanima Dutta

Unsupervised domain adaptation (UDA) for videos and 1D time-series data faces significant challenges due to domain shifts in terms of both temporal dynamics and feature distributions. Existing UDA approaches for time-series data often address temporal alignment and uncertainty mitigation as separate objectives, leading to unstable training, noisy pseudo-labels, and incomplete feature transfer. This disjoint treatment fails to capture inter-channel causal dependencies and also overlooks the impact of prediction uncertainty on adaptation quality. This limits the transferability of learned representations and results in suboptimal adaptation. To address the aforementioned limitations, we propose a novel UDA framework, named Causally-Regularized Optimal Transport (in short Causal-OT), that preserves domain-invariant causal mechanisms by embedding causal graph regularization into robust OT alignment process. First we estimate inter-channel causal graphs in both source and target domains and learn a transport plan that not only aligns feature distributions but also improves interpretability and minimizes the discrepancy between causal structures of the Granger graphs. However, pseudo-labeling may still prone to error propagation allowing incorrect target predictions during self-training, degrading the model stability and transfer quality across domains. To mitigate this, we further introduce a causality-aware pseudo-labeling strategy that selects high-confidence target samples based on both entropy and structural consistency with the causal graph of the source domain. This enhances robustness against pseudo-label noise.Extensive experiments on six time-series benchmarks achieving 4.5% gain in accuracy and a 3.8% improvement in F1-score. We conduct experiments on four benchmark video datasets that achieve a 2.5% gain in accuracy.

Farhat Shaikh, Ayan Banerjee, Sandeep Gupta

We introduce EMMA, a physics-informed multimodal framework that recovers all identifiable dynamical parameters of a system directly from raw video, audio, and image-based time-series observations. Unlike prior video-only approaches that struggle with occluded states, hidden actuation inputs, or assumptions about known initial conditions and coordinate frames, EMMA performs joint inference of explicit parameters, implicit dynamical components, and calibration invariants within a unified continuous-time model. EMMA leverages a Liquid Time-Constant (LTC) network to learn latent dynamics from heterogeneous modalities while a physics-constrained loss enforces consistency with the governing differential equations. A unified feature pipeline enables consistent alignment across video trajectories, acoustic signatures, and chart-derived measurements, allowing EMMA to estimate parameters under forced, implicit, and multivariate dynamics without requiring segmentation masks, differentiable rendering, or specialized sensors. Across 100+ scenarios including five standard dynamical benchmarks (75 Delfys videos), real-world rover and quadrotor systems with hidden inputs, and simulation-chart case studies spanning biological and chaotic systems, EMMA delivers robust multi-parameter recovery and significantly outperforms existing single-modality and equation-discovery baselines. Our results establish EMMA as a general, scalable solution for physics-consistent model extraction from opportunistic multimodal data.

Long Chen, Hui Wang, Man Xu, Zexuan Li, Zizhu Fan

The YOLO (You Only Look Once) series has been a cornerstone in real-time object detection, renowned for its efficient convolutional design and rapid inference. However, its reliance on convolutional operations inherently limits its ability to capture long-range dependencies and rich contextual information, leading to suboptimal performance in complex scenes. Recently, SSM (State Space Models) have emerged as an efficient alternative to attention mechanisms, offering global representation with linear time complexity. In this paper, we propose AKCMamba-YOLO, a novel object detector that incorporates SSM into the YOLO architecture. We introduce 3CAKCMamba and 4CAKCMamba modules to a novel object detection framework, enabling enhanced channel interaction and cross-layer semantic fusion. This design improves multi-scale feature modeling while maintaining computational efficiency. To support safety-critical applications, we provide railway pedestrian Detection datasets with 2,975 annotated images under complex scenarios. Experiments on COCO2017, power tower foreign object detection datasets, and our custom dataset show that AKCMamba-YOLO achieves superior accuracy and speed compared to state-of-the-art baselines, making it well-suited for real-time and resource-constrained environments.

Xingxing Xie, Jiahua Dong, Junwei Han, Gong Cheng

YOLO detectors are known for their fast inference speed, yet training them remains unexpectedly time-consuming due to their exhaustive pipeline that processes every training image in every epoch, even when many images have already been sufficiently learned. This stands in clear contrast to the efficiency suggested by the "You Only Look Once" philosophy. This naturally raises an important question: Does YOLO really need to see every training image in every epoch? To explore this, we propose an Anti-Forgetting Sampling Strategy (AFSS) that dynamically determines which images should be used and which can be skipped during each epoch, allowing the detector to learn more effectively and efficiently. Specifically, AFSS measures the learning sufficiency of each training image as the minimum of its detection recall and precision, and dynamically categorizes training images into easy, medium, or hard levels accordingly. Easy training images are sparsely resampled during training in a continuous review manner, with priority given to those that have not been used for a long time to reduce redundancy and prevent forgetting. Moderate training images are partially selected, prioritizing recently unused ones and randomly choosing the rest from unselected images to ensure coverage and prevent forgetting. Hard training images are fully sampled in every epoch to ensure sufficient learning. The learning sufficiency of each training image is periodically updated, enabling detectors to adaptively shift its focus toward the informative training images over time while progressively discarding redundant ones. On widely used natural image detection benchmarks (MS COCO 2017 and PASCAL VOC 2007) and remote sensing detection datasets (DOTA-v1.0 and DIOR-R), AFSS achieves more than 1.43x training speedup for YOLO-series detectors while also improving accuracy.