Xinlei Chen

Duy-Kien Nguyen, Vaibhav Aggarwal, Yanghao Li et al. · meta-ai

In this work, we explore regions as a potential visual analogue of words for self-supervised image representation learning. Inspired by Masked Autoencoding (MAE), a generative pre-training baseline, we propose masked region autoencoding to learn from groups of pixels or regions. Specifically, we design an architecture which efficiently addresses the one-to-many mapping between images and regions, while being highly effective especially with high-quality regions. When integrated with MAE, our approach (R-MAE) demonstrates consistent improvements across various pre-training datasets and downstream detection and segmentation benchmarks, with negligible computational overheads. Beyond the quantitative evaluation, our analysis indicates the models pre-trained with masked region autoencoding unlock the potential for interactive segmentation. The code is provided at https://github.com/facebookresearch/r-mae.

Yossi Gandelsman, Yu Sun, Xinlei Chen et al. · berkeley

Test-time training adapts to a new test distribution on the fly by optimizing a model for each test input using self-supervision. In this paper, we use masked autoencoders for this one-sample learning problem. Empirically, our simple method improves generalization on many visual benchmarks for distribution shifts. Theoretically, we characterize this improvement in terms of the bias-variance trade-off.

Renhao Wang, Yu Sun, Arnuv Tandon et al. · berkeley, ibm-research

Prior work has established Test-Time Training (TTT) as a general framework to further improve a trained model at test time. Before making a prediction on each test instance, the model is first trained on the same instance using a self-supervised task such as reconstruction. We extend TTT to the streaming setting, where multiple test instances - video frames in our case - arrive in temporal order. Our extension is online TTT: The current model is initialized from the previous model, then trained on the current frame and a small window of frames immediately before. Online TTT significantly outperforms the fixed-model baseline for four tasks, on three real-world datasets. The improvements are more than 2.2x and 1.5x for instance and panoptic segmentation. Surprisingly, online TTT also outperforms its offline variant that accesses strictly more information, training on all frames from the entire test video regardless of temporal order. This finding challenges those in prior work using synthetic videos. We formalize a notion of locality as the advantage of online over offline TTT, and analyze its role with ablations and a theory based on bias-variance trade-off.

Jie Lei, Xinlei Chen, Ning Zhang et al. · meta-ai

Dual encoders and cross encoders have been widely used for image-text retrieval. Between the two, the dual encoder encodes the image and text independently followed by a dot product, while the cross encoder jointly feeds image and text as the input and performs dense multi-modal fusion. These two architectures are typically modeled separately without interaction. In this work, we propose LoopITR, which combines them in the same network for joint learning. Specifically, we let the dual encoder provide hard negatives to the cross encoder, and use the more discriminative cross encoder to distill its predictions back to the dual encoder. Both steps are efficiently performed together in the same model. Our work centers on empirical analyses of this combined architecture, putting the main focus on the design of the distillation objective. Our experimental results highlight the benefits of training the two encoders in the same network, and demonstrate that distillation can be quite effective with just a few hard negative examples. Experiments on two standard datasets (Flickr30K and COCO) show our approach achieves state-of-the-art dual encoder performance when compared with approaches using a similar amount of data.

Sanghyun Woo, Shoubhik Debnath, Ronghang Hu et al.

Driven by improved architectures and better representation learning frameworks, the field of visual recognition has enjoyed rapid modernization and performance boost in the early 2020s. For example, modern ConvNets, represented by ConvNeXt, have demonstrated strong performance in various scenarios. While these models were originally designed for supervised learning with ImageNet labels, they can also potentially benefit from self-supervised learning techniques such as masked autoencoders (MAE). However, we found that simply combining these two approaches leads to subpar performance. In this paper, we propose a fully convolutional masked autoencoder framework and a new Global Response Normalization (GRN) layer that can be added to the ConvNeXt architecture to enhance inter-channel feature competition. This co-design of self-supervised learning techniques and architectural improvement results in a new model family called ConvNeXt V2, which significantly improves the performance of pure ConvNets on various recognition benchmarks, including ImageNet classification, COCO detection, and ADE20K segmentation. We also provide pre-trained ConvNeXt V2 models of various sizes, ranging from an efficient 3.7M-parameter Atto model with 76.7% top-1 accuracy on ImageNet, to a 650M Huge model that achieves a state-of-the-art 88.9% accuracy using only public training data.

Dave Zhenyu Chen, Ronghang Hu, Xinlei Chen et al.

Performing 3D dense captioning and visual grounding requires a common and shared understanding of the underlying multimodal relationships. However, despite some previous attempts on connecting these two related tasks with highly task-specific neural modules, it remains understudied how to explicitly depict their shared nature to learn them simultaneously. In this work, we propose UniT3D, a simple yet effective fully unified transformer-based architecture for jointly solving 3D visual grounding and dense captioning. UniT3D enables learning a strong multimodal representation across the two tasks through a supervised joint pre-training scheme with bidirectional and seq-to-seq objectives. With a generic architecture design, UniT3D allows expanding the pre-training scope to more various training sources such as the synthesized data from 2D prior knowledge to benefit 3D vision-language tasks. Extensive experiments and analysis demonstrate that UniT3D obtains significant gains for 3D dense captioning and visual grounding.

Xiao Wang, Haoqi Fan, Yuandong Tian et al.

Many recent self-supervised frameworks for visual representation learning are based on certain forms of Siamese networks. Such networks are conceptually symmetric with two parallel encoders, but often practically asymmetric as numerous mechanisms are devised to break the symmetry. In this work, we conduct a formal study on the importance of asymmetry by explicitly distinguishing the two encoders within the network -- one produces source encodings and the other targets. Our key insight is keeping a relatively lower variance in target than source generally benefits learning. This is empirically justified by our results from five case studies covering different variance-oriented designs, and is aligned with our preliminary theoretical analysis on the baseline. Moreover, we find the improvements from asymmetric designs generalize well to longer training schedules, multiple other frameworks and newer backbones. Finally, the combined effect of several asymmetric designs achieves a state-of-the-art accuracy on ImageNet linear probing and competitive results on downstream transfer. We hope our exploration will inspire more research in exploiting asymmetry for Siamese representation learning.

Nan Zhou, Huandong Wang, Jiahao Li et al.

Fine-grained wildfire spread prediction is crucial for enhancing emergency response efficacy and decision-making precision. However, existing research predominantly focuses on coarse spatiotemporal scales and relies on low-resolution satellite data, capturing only macroscopic fire states while fundamentally constraining high-precision localized fire dynamics modeling capabilities. To bridge this gap, we present FireSentry, a provincial-scale multi-modal wildfire dataset characterized by sub-meter spatial and sub-second temporal resolution. Collected using synchronized UAV platforms, FireSentry provides visible and infrared video streams, in-situ environmental measurements, and manually validated fire masks. Building on FireSentry, we establish a comprehensive benchmark encompassing physics-based, data-driven, and generative models, revealing the limitations of existing mask-only approaches. Our analysis proposes FiReDiff, a novel dual-modality paradigm that first predicts future video sequences in the infrared modality, and then precisely segments fire masks in the mask modality based on the generated dynamics. FiReDiff achieves state-of-the-art performance, with video quality gains of 39.2% in PSNR, 36.1% in SSIM, 50.0% in LPIPS, 29.4% in FVD, and mask accuracy gains of 3.3% in AUPRC, 59.1% in F1 score, 42.9% in IoU, and 62.5% in MSE when applied to generative models. The FireSentry benchmark dataset and FiReDiff paradigm collectively advance fine-grained wildfire forecasting and dynamic disaster simulation. The processed benchmark dataset is publicly available at: https://github.com/Munan222/FireSentry-Benchmark-Dataset.

Ronghang Hu, Shoubhik Debnath, Saining Xie et al.

Masked Autoencoding (MAE) has emerged as an effective approach for pre-training representations across multiple domains. In contrast to discrete tokens in natural languages, the input for image MAE is continuous and subject to additional specifications. We systematically study each input specification during the pre-training stage, and find sequence length is a key axis that further scales MAE. Our study leads to a long-sequence version of MAE with minimal changes to the original recipe, by just decoupling the mask size from the patch size. For object detection and semantic segmentation, our long-sequence MAE shows consistent gains across all the experimental setups without extra computation cost during the transfer. While long-sequence pre-training is discerned most beneficial for detection and segmentation, we also achieve strong results on ImageNet-1K classification by keeping a standard image size and only increasing the sequence length. We hope our findings can provide new insights and avenues for scaling in computer vision.

Yu Sun, Xinhao Li, Karan Dalal et al.

We reformulate the problem of supervised learning as learning to learn with two nested loops (i.e. learning problems). The inner loop learns on each individual instance with self-supervision before final prediction. The outer loop learns the self-supervised task used by the inner loop, such that its final prediction improves. Our inner loop turns out to be equivalent to linear attention when the inner-loop learner is only a linear model, and to self-attention when it is a kernel estimator. For practical comparison with linear or self-attention layers, we replace each of them in a transformer with an inner loop, so our outer loop is equivalent to training the architecture. When each inner-loop learner is a neural network, our approach vastly outperforms transformers with linear attention on ImageNet from 224 x 224 raw pixels in both accuracy and FLOPs, while (regular) transformers cannot run.

Xinqi Jin, Zhui Zhu, Xikai Sun et al.

Neural enhancement through super-resolution (SR) deep neural networks (DNNs) opens up new possibilities for ultra-high-definition (UHD) live streaming over existing encoding and networking infrastructure. Yet, the heavy SR DNN inference overhead leads to severe deployment challenges. To reduce the overhead, existing systems propose to apply DNN-based SR only on carefully selected anchor frames while upscaling non-anchor frames via the lightweight reusing-based SR approach. However, frame-level scheduling is coarse-grained and fails to deliver optimal efficiency. In this work, we propose Palantir, the first neural-enhanced UHD live streaming system with fine-grained patch-level scheduling. Two novel techniques are incorporated into Palantir to select the most beneficial anchor patches and support latency-sensitive UHD live streaming applications. Firstly, under the guidance of our pioneering and theoretical analysis, Palantir constructs a directed acyclic graph (DAG) for lightweight yet accurate SR quality estimation under any possible anchor patch set. Secondly, to further optimize the scheduling latency, Palantir improves parallelizability by refactoring the computation subprocedure of the estimation process into a sparse matrix-matrix multiplication operation. The evaluation results suggest that Palantir incurs a negligible scheduling latency accounting for less than 5.7% of the end-to-end latency requirement. When compared to the naive method of applying DNN-based SR on all the frames, Palantir can reduce the SR DNN inference overhead by 20 times (or 60 times) while preserving 54.0-82.6% (or 32.8-64.0%) of the quality gain. When compared to the state-of-the-art real-time frame-level scheduling strategy, Palantir can reduce the SR DNN inference overhead by 80.1% at most (and 38.4% on average) without sacrificing the video quality.

Haoyang Wang, Ruishan Guo, Pengtao Ma et al.

With the increasing complexity of mobile device applications, these devices are evolving toward high agility. This shift imposes new demands on mobile sensing, particularly in achieving high-accuracy and low-latency. Event-based vision has emerged as a disruptive paradigm, offering high temporal resolution and low latency, making it well-suited for high-accuracy and low-latency sensing tasks on high-agility platforms. However, the presence of substantial noisy events, lack of stable, persistent semantic information, and large data volume pose challenges for event-based data processing on resource-constrained mobile devices. This paper surveys the literature from 2014 to 2025 and presents a comprehensive overview of event-based mobile sensing, encompassing its fundamental principles, event \textit{abstraction} methods, \textit{algorithm} advancements, and both hardware and software \textit{acceleration} strategies. We discuss key \textit{applications} of event cameras in mobile sensing, including visual odometry, object tracking, optical flow, and 3D reconstruction, while highlighting challenges associated with event data processing, sensor fusion, and real-time deployment. Furthermore, we outline future research directions, such as improving the event camera with advanced optics, leveraging neuromorphic computing for efficient processing, and integrating bio-inspired algorithms. To support ongoing research, we provide an open-source \textit{Online Sheet} with recent developments. We hope this survey serves as a reference, facilitating the adoption of event-based vision across diverse applications.

Kaiyuan Li, Xiaoyue Chen, Chen Gao et al.

Large Vision-Language Models (LVLMs) have shown impressive performance across multi-modal tasks by encoding images into thousands of tokens. However, the large number of image tokens results in significant computational overhead, and the use of dynamic high-resolution inputs further increases this burden. Previous approaches have attempted to reduce the number of image tokens through token pruning, typically by selecting tokens based on attention scores or image token diversity. Through empirical studies, we observe that existing methods often overlook the joint impact of pruning on both the current layer's output (local) and the outputs of subsequent layers (global), leading to suboptimal pruning decisions. To address this challenge, we propose Balanced Token Pruning (BTP), a plug-and-play method for pruning vision tokens. Specifically, our method utilizes a small calibration set to divide the pruning process into multiple stages. In the early stages, our method emphasizes the impact of pruning on subsequent layers, whereas in the deeper stages, the focus shifts toward preserving the consistency of local outputs. Extensive experiments across various LVLMs demonstrate the broad effectiveness of our approach on multiple benchmarks. Our method achieves a 78% compression rate while preserving 96.7% of the original models' performance on average. Our code is available at https://github.com/EmbodiedCity/NeurIPS2025-Balanced-Token-Pruning.

Adrien Bardes, Quentin Garrido, Jean Ponce et al.

This paper explores feature prediction as a stand-alone objective for unsupervised learning from video and introduces V-JEPA, a collection of vision models trained solely using a feature prediction objective, without the use of pretrained image encoders, text, negative examples, reconstruction, or other sources of supervision. The models are trained on 2 million videos collected from public datasets and are evaluated on downstream image and video tasks. Our results show that learning by predicting video features leads to versatile visual representations that perform well on both motion and appearance-based tasks, without adaption of the model's parameters; e.g., using a frozen backbone. Our largest model, a ViT-H/16 trained only on videos, obtains 81.9% on Kinetics-400, 72.2% on Something-Something-v2, and 77.9% on ImageNet1K.

Shengbang Tong, David Fan, Jiachen Zhu et al.

In this work, we propose Visual-Predictive Instruction Tuning (VPiT) - a simple and effective extension to visual instruction tuning that enables a pretrained LLM to quickly morph into an unified autoregressive model capable of generating both text and visual tokens. VPiT teaches an LLM to predict discrete text tokens and continuous visual tokens from any input sequence of image and text data curated in an instruction-following format. Our empirical investigation reveals several intriguing properties of VPiT: (1) visual generation ability emerges as a natural byproduct of improved visual understanding, and can be unlocked efficiently with a small amount of generation data; (2) while we find understanding and generation to be mutually beneficial, understanding data contributes to both capabilities more effectively than generation data. Building upon these findings, we train our MetaMorph model and achieve competitive performance on both visual understanding and generation. In visual generation, MetaMorph can leverage the world knowledge and reasoning abilities gained from LLM pretraining, and overcome common failure modes exhibited by other generation models. Our results suggest that LLMs may have strong "prior" vision capabilities that can be efficiently adapted to both visual understanding and generation with a relatively simple instruction tuning process.

Jirong Zha, Yuxuan Fan, Tianyu Zhang et al.

Multimodal Large Language Models (MLLMs) have shown promise in single-agent vision tasks, yet benchmarks for evaluating multi-agent collaborative perception remain scarce. This gap is critical, as multi-drone systems provide enhanced coverage, robustness, and collaboration compared to single-sensor setups. Existing multi-image benchmarks mainly target basic perception tasks using high-quality single-agent images, thus failing to evaluate MLLMs in more complex, egocentric collaborative scenarios, especially under real-world degraded perception conditions.To address these challenges, we introduce AirCopBench, the first comprehensive benchmark designed to evaluate MLLMs in embodied aerial collaborative perception under challenging perceptual conditions. AirCopBench includes 14.6k+ questions derived from both simulator and real-world data, spanning four key task dimensions: Scene Understanding, Object Understanding, Perception Assessment, and Collaborative Decision, across 14 task types. We construct the benchmark using data from challenging degraded-perception scenarios with annotated collaborative events, generating large-scale questions through model-, rule-, and human-based methods under rigorous quality control. Evaluations on 40 MLLMs show significant performance gaps in collaborative perception tasks, with the best model trailing humans by 24.38% on average and exhibiting inconsistent results across tasks. Fine-tuning experiments further confirm the feasibility of sim-to-real transfer in aerial collaborative perception and reasoning.

Yu Jiang, Vivek Natarajan, Xinlei Chen et al.

This document describes Pythia v0.1, the winning entry from Facebook AI Research (FAIR)'s A-STAR team to the VQA Challenge 2018. Our starting point is a modular re-implementation of the bottom-up top-down (up-down) model. We demonstrate that by making subtle but important changes to the model architecture and the learning rate schedule, fine-tuning image features, and adding data augmentation, we can significantly improve the performance of the up-down model on VQA v2.0 dataset -- from 65.67% to 70.22%. Furthermore, by using a diverse ensemble of models trained with different features and on different datasets, we are able to significantly improve over the 'standard' way of ensembling (i.e. same model with different random seeds) by 1.31%. Overall, we achieve 72.27% on the test-std split of the VQA v2.0 dataset. Our code in its entirety (training, evaluation, data-augmentation, ensembling) and pre-trained models are publicly available at: https://github.com/facebookresearch/pythia

Jiachen Zhu, Xinlei Chen, Kaiming He et al.

Normalization layers are ubiquitous in modern neural networks and have long been considered essential. This work demonstrates that Transformers without normalization can achieve the same or better performance using a remarkably simple technique. We introduce Dynamic Tanh (DyT), an element-wise operation $DyT($x$) = \tanh(α$x$)$, as a drop-in replacement for normalization layers in Transformers. DyT is inspired by the observation that layer normalization in Transformers often produces tanh-like, $S$-shaped input-output mappings. By incorporating DyT, Transformers without normalization can match or exceed the performance of their normalized counterparts, mostly without hyperparameter tuning. We validate the effectiveness of Transformers with DyT across diverse settings, ranging from recognition to generation, supervised to self-supervised learning, and computer vision to language models. These findings challenge the conventional understanding that normalization layers are indispensable in modern neural networks, and offer new insights into their role in deep networks.

Chen Gao, Baining Zhao, Weichen Zhang et al.

Embodied artificial intelligence emphasizes the role of an agent's body in generating human-like behaviors. The recent efforts on EmbodiedAI pay a lot of attention to building up machine learning models to possess perceiving, planning, and acting abilities, thereby enabling real-time interaction with the world. However, most works focus on bounded indoor environments, such as navigation in a room or manipulating a device, with limited exploration of embodying the agents in open-world scenarios. That is, embodied intelligence in the open and outdoor environment is less explored, for which one potential reason is the lack of high-quality simulators, benchmarks, and datasets. To address it, in this paper, we construct a benchmark platform for embodied intelligence evaluation in real-world city environments. Specifically, we first construct a highly realistic 3D simulation environment based on the real buildings, roads, and other elements in a real city. In this environment, we combine historically collected data and simulation algorithms to conduct simulations of pedestrian and vehicle flows with high fidelity. Further, we designed a set of evaluation tasks covering different EmbodiedAI abilities. Moreover, we provide a complete set of input and output interfaces for access, enabling embodied agents to easily take task requirements and current environmental observations as input and then make decisions and obtain performance evaluations. On the one hand, it expands the capability of existing embodied intelligence to higher levels. On the other hand, it has a higher practical value in the real world and can support more potential applications for artificial general intelligence. Based on this platform, we evaluate some popular large language models for embodied intelligence capabilities of different dimensions and difficulties.

Baining Zhao, Ziyou Wang, Jianjie Fang et al.

Humans can perceive and reason about spatial relationships from sequential visual observations, such as egocentric video streams. However, how pretrained models acquire such abilities, especially high-level reasoning, remains unclear. This paper introduces Embodied-R, a collaborative framework combining large-scale Vision-Language Models (VLMs) for perception and small-scale Language Models (LMs) for reasoning. Using Reinforcement Learning (RL) with a novel reward system considering think-answer logical consistency, the model achieves slow-thinking capabilities with limited computational resources. After training on only 5k embodied video samples, Embodied-R with a 3B LM matches state-of-the-art multimodal reasoning models (OpenAI-o1, Gemini-2.5-pro) on both in-distribution and out-of-distribution embodied spatial reasoning tasks. Embodied-R also exhibits emergent thinking patterns such as systematic analysis and contextual integration. We further explore research questions including response length, training on VLM, strategies for reward design, and differences in model generalization after SFT (Supervised Fine-Tuning) and RL training.

Baining Zhao, Jianjie Fang, Zichao Dai et al.

Large multimodal models exhibit remarkable intelligence, yet their embodied cognitive abilities during motion in open-ended urban 3D space remain to be explored. We introduce a benchmark to evaluate whether video-large language models (Video-LLMs) can naturally process continuous first-person visual observations like humans, enabling recall, perception, reasoning, and navigation. We have manually control drones to collect 3D embodied motion video data from real-world cities and simulated environments, resulting in 1.5k video clips. Then we design a pipeline to generate 5.2k multiple-choice questions. Evaluations of 17 widely-used Video-LLMs reveal current limitations in urban embodied cognition. Correlation analysis provides insight into the relationships between different tasks, showing that causal reasoning has a strong correlation with recall, perception, and navigation, while the abilities for counterfactual and associative reasoning exhibit lower correlation with other tasks. We also validate the potential for Sim-to-Real transfer in urban embodiment through fine-tuning.

Philippe Hansen-Estruch, David Yan, Ching-Yao Chung et al.

Visual tokenization via auto-encoding empowers state-of-the-art image and video generative models by compressing pixels into a latent space. Although scaling Transformer-based generators has been central to recent advances, the tokenizer component itself is rarely scaled, leaving open questions about how auto-encoder design choices influence both its objective of reconstruction and downstream generative performance. Our work aims to conduct an exploration of scaling in auto-encoders to fill in this blank. To facilitate this exploration, we replace the typical convolutional backbone with an enhanced Vision Transformer architecture for Tokenization (ViTok). We train ViTok on large-scale image and video datasets far exceeding ImageNet-1K, removing data constraints on tokenizer scaling. We first study how scaling the auto-encoder bottleneck affects both reconstruction and generation -- and find that while it is highly correlated with reconstruction, its relationship with generation is more complex. We next explored the effect of separately scaling the auto-encoders' encoder and decoder on reconstruction and generation performance. Crucially, we find that scaling the encoder yields minimal gains for either reconstruction or generation, while scaling the decoder boosts reconstruction but the benefits for generation are mixed. Building on our exploration, we design ViTok as a lightweight auto-encoder that achieves competitive performance with state-of-the-art auto-encoders on ImageNet-1K and COCO reconstruction tasks (256p and 512p) while outperforming existing auto-encoders on 16-frame 128p video reconstruction for UCF-101, all with 2-5x fewer FLOPs. When integrated with Diffusion Transformers, ViTok demonstrates competitive performance on image generation for ImageNet-1K and sets new state-of-the-art benchmarks for class-conditional video generation on UCF-101.

Haoyang Wang, Jingao Xu, Xinyu Luo et al.

For precise, efficient, and safe drone landings, ground platforms should real-time, accurately locate descending drones and guide them to designated spots. While mmWave sensing combined with cameras improves localization accuracy, the lower sampling frequency of traditional frame cameras compared to mmWave radar creates bottlenecks in system throughput. In this work, we replace the traditional frame camera with event camera, a novel sensor that harmonizes in sampling frequency with mmWave radar within the ground platform setup, and introduce mmE-Loc, a high-precision, low-latency ground localization system designed for drone landings. To fully leverage the \textit{temporal consistency} and \textit{spatial complementarity} between these modalities, we propose two innovative modules, \textit{consistency-instructed collaborative tracking} and \textit{graph-informed adaptive joint optimization}, for accurate drone measurement extraction and efficient sensor fusion. Extensive real-world experiments in landing scenarios from a leading drone delivery company demonstrate that mmE-Loc outperforms state-of-the-art methods in both localization accuracy and latency.

Lirui Wang, Kevin Zhao, Chaoqi Liu et al.

We propose Heterogeneous Masked Autoregression (HMA) for modeling action-video dynamics to generate high-quality data and evaluation in scaling robot learning. Building interactive video world models and policies for robotics is difficult due to the challenge of handling diverse settings while maintaining computational efficiency to run in real time. HMA uses heterogeneous pre-training from observations and action sequences across different robotic embodiments, domains, and tasks. HMA uses masked autoregression to generate quantized or soft tokens for video predictions. \ourshort achieves better visual fidelity and controllability than the previous robotic video generation models with 15 times faster speed in the real world. After post-training, this model can be used as a video simulator from low-level action inputs for evaluating policies and generating synthetic data. See this link https://liruiw.github.io/hma for more information.

Alexander C. Li, Yuandong Tian, Beidi Chen et al.

Conventional wisdom suggests that pre-training Vision Transformers (ViT) improves downstream performance by learning useful representations. Is this actually true? We investigate this question and find that the features and representations learned during pre-training are not essential. Surprisingly, using only the attention patterns from pre-training (i.e., guiding how information flows between tokens) is sufficient for models to learn high quality features from scratch and achieve comparable downstream performance. We show this by introducing a simple method called attention transfer, where only the attention patterns from a pre-trained teacher ViT are transferred to a student, either by copying or distilling the attention maps. Since attention transfer lets the student learn its own features, ensembling it with a fine-tuned teacher also further improves accuracy on ImageNet. We systematically study various aspects of our findings on the sufficiency of attention maps, including distribution shift settings where they underperform fine-tuning. We hope our exploration provides a better understanding of what pre-training accomplishes and leads to a useful alternative to the standard practice of fine-tuning

Yung-Sung Chuang, Yang Li, Dong Wang et al. · meta-ai, mit

Contrastive Language-Image Pretraining (CLIP) is a popular foundation model, supporting from zero-shot classification, retrieval to encoders for multimodal large language models (MLLMs). Although CLIP is successfully trained on billion-scale image-text pairs from the English world, scaling CLIP's training further to learning from the worldwide web data is still challenging: (1) no curation method is available to handle data points from non-English world; (2) the English performance from existing multilingual CLIP is worse than its English-only counterpart, i.e., "curse of multilinguality" that is common in LLMs. Here, we present Meta CLIP 2, the first recipe training CLIP from scratch on worldwide web-scale image-text pairs. To generalize our findings, we conduct rigorous ablations with minimal changes that are necessary to address the above challenges and present a recipe enabling mutual benefits from English and non-English world data. In zero-shot ImageNet classification, Meta CLIP 2 ViT-H/14 surpasses its English-only counterpart by 0.8% and mSigLIP by 0.7%, and surprisingly sets new state-of-the-art without system-level confounding factors (e.g., translation, bespoke architecture changes) on multilingual benchmarks, such as CVQA with 57.4%, Babel-ImageNet with 50.2% and XM3600 with 64.3% on image-to-text retrieval.

Ciyu Ruan, Ruishan Guo, Zihang Gong et al.

Event cameras excel in high temporal resolution and dynamic range but suffer from dense noise in rainy conditions. Existing event deraining methods face trade-offs between temporal precision, deraining effectiveness, and computational efficiency. In this paper, we propose PRE-Mamba, a novel point-based event camera deraining framework that fully exploits the spatiotemporal characteristics of raw event and rain. Our framework introduces a 4D event cloud representation that integrates dual temporal scales to preserve high temporal precision, a Spatio-Temporal Decoupling and Fusion module (STDF) that enhances deraining capability by enabling shallow decoupling and interaction of temporal and spatial information, and a Multi-Scale State Space Model (MS3M) that captures deeper rain dynamics across dual-temporal and multi-spatial scales with linear computational complexity. Enhanced by frequency-domain regularization, PRE-Mamba achieves superior performance (0.95 SR, 0.91 NR, and 0.4s/M events) with only 0.26M parameters on EventRain-27K, a comprehensive dataset with labeled synthetic and real-world sequences. Moreover, our method generalizes well across varying rain intensities, viewpoints, and even snowy conditions.

Kumar Ashutosh, Yossi Gandelsman, Xinlei Chen et al. · meta-ai

We present MILS: Multimodal Iterative LLM Solver, a surprisingly simple, training-free approach, to imbue multimodal capabilities into your favorite LLM. Leveraging their innate ability to perform multi-step reasoning, MILS prompts the LLM to generate candidate outputs, each of which are scored and fed back iteratively, eventually generating a solution to the task. This enables various applications that typically require training specialized models on task-specific data. In particular, we establish a new state-of-the-art on emergent zero-shot image, video and audio captioning. MILS seamlessly applies to media generation as well, discovering prompt rewrites to improve text-to-image generation, and even edit prompts for style transfer! Finally, being a gradient-free optimization approach, MILS can invert multimodal embeddings into text, enabling applications like cross-modal arithmetic.

Weichen Zhang, Ruiying Peng, Chen Gao et al.

3D Large Language Models (LLMs) leveraging spatial information in point clouds for 3D spatial reasoning attract great attention. Despite some promising results, the role of point clouds in 3D spatial reasoning remains under-explored. In this work, we comprehensively evaluate and analyze these models to answer the research question: \textit{Does point cloud truly boost the spatial reasoning capacities of 3D LLMs?} We first evaluate the spatial reasoning capacity of LLMs with different input modalities by replacing the point cloud with the visual and text counterparts. We then propose a novel 3D QA (Question-answering) benchmark, ScanReQA, that comprehensively evaluates models' understanding of binary spatial relationships. Our findings reveal several critical insights: 1) LLMs without point input could even achieve competitive performance even in a zero-shot manner; 2) existing 3D LLMs struggle to comprehend the binary spatial relationships; 3) 3D LLMs exhibit limitations in exploiting the structural coordinates in point clouds for fine-grained spatial reasoning. We think these conclusions can help the next step of 3D LLMs and also offer insights for foundation models in other modalities. We release datasets and reproducible codes in the anonymous project page: https://3d-llm.xyz.

Ruiying Peng, Kaiyuan Li, Weichen Zhang et al.

Recently, 3D-LLMs, which combine point-cloud encoders with large models, have been proposed to tackle complex tasks in embodied intelligence and scene understanding. In addition to showing promising results on 3D tasks, we found that they are significantly affected by hallucinations. For instance, they may generate objects that do not exist in the scene or produce incorrect relationships between objects. To investigate this issue, this work presents the first systematic study of hallucinations in 3D-LLMs. We begin by quickly evaluating hallucinations in several representative 3D-LLMs and reveal that they are all significantly affected by hallucinations. We then define hallucinations in 3D scenes and, through a detailed analysis of datasets, uncover the underlying causes of these hallucinations. We find three main causes: (1) Uneven frequency distribution of objects in the dataset. (2) Strong correlations between objects. (3) Limited diversity in object attributes. Additionally, we propose new evaluation metrics for hallucinations, including Random Point Cloud Pair and Opposite Question Evaluations, to assess whether the model generates responses based on visual information and aligns it with the text's meaning.

Ciyu Ruan, Zihang Gong, Ruishan Guo et al.

Event cameras provide micro-second latency and broad dynamic range, yet their raw streams are marred by spatial artifacts (e.g., hot pixels) and temporally inconsistent background activity. Existing methods jointly process the entire 4D event volume (x, y, p, t), forcing heavy spatio-temporal attention that inflates parameters, FLOPs, and latency. We introduce EDmamba, a compact event-denoising framework that embraces the key insight that spatial and temporal noise arise from different physical mechanisms and can therefore be suppressed independently. A polarity- and geometry-aware encoder first extracts coarse cues, which are then routed to two lightweight state-space branches: a Spatial-SSM that learns location-conditioned filters to silence persistent artifacts, and a Temporal-SSM that models causal signal dynamics to eliminate bursty background events. This decoupled design distills the network to only 88.9K parameters and 2.27GFLOPs, enabling real-time throughput of 100K events in 68ms on a single GPU, 36x faster than recent Transformer baselines. Despite its economy, EDmamba establishes new state-of-the-art accuracy on four public benchmarks, outscoring the strongest prior model by 2.1 percentage points.

Duy-Kien Nguyen, Mahmoud Assran, Unnat Jain et al.

This work does not introduce a new method. Instead, we present an interesting finding that questions the necessity of the inductive bias of locality in modern computer vision architectures. Concretely, we find that vanilla Transformers can operate by directly treating each individual pixel as a token and achieve highly performant results. This is substantially different from the popular design in Vision Transformer, which maintains the inductive bias from ConvNets towards local neighborhoods (e.g. by treating each 16x16 patch as a token). We showcase the effectiveness of pixels-as-tokens across three well-studied computer vision tasks: supervised learning for classification and regression, self-supervised learning via masked autoencoding, and image generation with diffusion models. Although it's computationally less practical to directly operate on individual pixels, we believe the community must be made aware of this surprising piece of knowledge when devising the next generation of neural network architectures for computer vision.

Xinlei Chen, Zhuang Liu, Saining Xie et al.

In this study, we examine the representation learning abilities of Denoising Diffusion Models (DDM) that were originally purposed for image generation. Our philosophy is to deconstruct a DDM, gradually transforming it into a classical Denoising Autoencoder (DAE). This deconstructive procedure allows us to explore how various components of modern DDMs influence self-supervised representation learning. We observe that only a very few modern components are critical for learning good representations, while many others are nonessential. Our study ultimately arrives at an approach that is highly simplified and to a large extent resembles a classical DAE. We hope our study will rekindle interest in a family of classical methods within the realm of modern self-supervised learning.

Yutong Bai, Xinlei Chen, Alexander Kirillov et al.

In this work we present point-level region contrast, a self-supervised pre-training approach for the task of object detection. This approach is motivated by the two key factors in detection: localization and recognition. While accurate localization favors models that operate at the pixel- or point-level, correct recognition typically relies on a more holistic, region-level view of objects. Incorporating this perspective in pre-training, our approach performs contrastive learning by directly sampling individual point pairs from different regions. Compared to an aggregated representation per region, our approach is more robust to the change in input region quality, and further enables us to implicitly improve initial region assignments via online knowledge distillation during training. Both advantages are important when dealing with imperfect regions encountered in the unsupervised setting. Experiments show point-level region contrast improves on state-of-the-art pre-training methods for object detection and segmentation across multiple tasks and datasets, and we provide extensive ablation studies and visualizations to aid understanding. Code will be made available.

Xinlei Chen, Saining Xie, Kaiming He

This paper does not describe a novel method. Instead, it studies a straightforward, incremental, yet must-know baseline given the recent progress in computer vision: self-supervised learning for Vision Transformers (ViT). While the training recipes for standard convolutional networks have been highly mature and robust, the recipes for ViT are yet to be built, especially in the self-supervised scenarios where training becomes more challenging. In this work, we go back to basics and investigate the effects of several fundamental components for training self-supervised ViT. We observe that instability is a major issue that degrades accuracy, and it can be hidden by apparently good results. We reveal that these results are indeed partial failure, and they can be improved when training is made more stable. We benchmark ViT results in MoCo v3 and several other self-supervised frameworks, with ablations in various aspects. We discuss the currently positive evidence as well as challenges and open questions. We hope that this work will provide useful data points and experience for future research.

Xinlei Chen, Kaiming He

Siamese networks have become a common structure in various recent models for unsupervised visual representation learning. These models maximize the similarity between two augmentations of one image, subject to certain conditions for avoiding collapsing solutions. In this paper, we report surprising empirical results that simple Siamese networks can learn meaningful representations even using none of the following: (i) negative sample pairs, (ii) large batches, (iii) momentum encoders. Our experiments show that collapsing solutions do exist for the loss and structure, but a stop-gradient operation plays an essential role in preventing collapsing. We provide a hypothesis on the implication of stop-gradient, and further show proof-of-concept experiments verifying it. Our "SimSiam" method achieves competitive results on ImageNet and downstream tasks. We hope this simple baseline will motivate people to rethink the roles of Siamese architectures for unsupervised representation learning. Code will be made available.

Medhini Narasimhan, Erik Wijmans, Xinlei Chen et al.

We introduce a learning-based approach for room navigation using semantic maps. Our proposed architecture learns to predict top-down belief maps of regions that lie beyond the agent's field of view while modeling architectural and stylistic regularities in houses. First, we train a model to generate amodal semantic top-down maps indicating beliefs of location, size, and shape of rooms by learning the underlying architectural patterns in houses. Next, we use these maps to predict a point that lies in the target room and train a policy to navigate to the point. We empirically demonstrate that by predicting semantic maps, the model learns common correlations found in houses and generalizes to novel environments. We also demonstrate that reducing the task of room navigation to point navigation improves the performance further.

Duy-Kien Nguyen, Vedanuj Goswami, Xinlei Chen

This paper focuses on visual counting, which aims to predict the number of occurrences given a natural image and a query (e.g. a question or a category). Unlike most prior works that use explicit, symbolic models which can be computationally expensive and limited in generalization, we propose a simple and effective alternative by revisiting modulated convolutions that fuse the query and the image locally. Following the design of residual bottleneck, we call our method MoVie, short for Modulated conVolutional bottlenecks. Notably, MoVie reasons implicitly and holistically and only needs a single forward-pass during inference. Nevertheless, MoVie showcases strong performance for counting: 1) advancing the state-of-the-art on counting-specific VQA tasks while being more efficient; 2) outperforming prior-art on difficult benchmarks like COCO for common object counting; 3) helped us secure the first place of 2020 VQA challenge when integrated as a module for 'number' related questions in generic VQA models. Finally, we show evidence that modulated convolutions such as MoVie can serve as a general mechanism for reasoning tasks beyond counting.

Charles R. Qi, Xinlei Chen, Or Litany et al.

3D object detection has seen quick progress thanks to advances in deep learning on point clouds. A few recent works have even shown state-of-the-art performance with just point clouds input (e.g. VoteNet). However, point cloud data have inherent limitations. They are sparse, lack color information and often suffer from sensor noise. Images, on the other hand, have high resolution and rich texture. Thus they can complement the 3D geometry provided by point clouds. Yet how to effectively use image information to assist point cloud based detection is still an open question. In this work, we build on top of VoteNet and propose a 3D detection architecture called ImVoteNet specialized for RGB-D scenes. ImVoteNet is based on fusing 2D votes in images and 3D votes in point clouds. Compared to prior work on multi-modal detection, we explicitly extract both geometric and semantic features from the 2D images. We leverage camera parameters to lift these features to 3D. To improve the synergy of 2D-3D feature fusion, we also propose a multi-tower training scheme. We validate our model on the challenging SUN RGB-D dataset, advancing state-of-the-art results by 5.7 mAP. We also provide rich ablation studies to analyze the contribution of each design choice.

Huaizu Jiang, Ishan Misra, Marcus Rohrbach et al.

Popularized as 'bottom-up' attention, bounding box (or region) based visual features have recently surpassed vanilla grid-based convolutional features as the de facto standard for vision and language tasks like visual question answering (VQA). However, it is not clear whether the advantages of regions (e.g. better localization) are the key reasons for the success of bottom-up attention. In this paper, we revisit grid features for VQA, and find they can work surprisingly well - running more than an order of magnitude faster with the same accuracy (e.g. if pre-trained in a similar fashion). Through extensive experiments, we verify that this observation holds true across different VQA models (reporting a state-of-the-art accuracy on VQA 2.0 test-std, 72.71), datasets, and generalizes well to other tasks like image captioning. As grid features make the model design and training process much simpler, this enables us to train them end-to-end and also use a more flexible network design. We learn VQA models end-to-end, from pixels directly to answers, and show that strong performance is achievable without using any region annotations in pre-training. We hope our findings help further improve the scientific understanding and the practical application of VQA. Code and features will be made available.

Licheng Yu, Xinlei Chen, Georgia Gkioxari et al.

Embodied Question Answering (EQA) is a relatively new task where an agent is asked to answer questions about its environment from egocentric perception. EQA makes the fundamental assumption that every question, e.g., "what color is the car?", has exactly one target ("car") being inquired about. This assumption puts a direct limitation on the abilities of the agent. We present a generalization of EQA - Multi-Target EQA (MT-EQA). Specifically, we study questions that have multiple targets in them, such as "Is the dresser in the bedroom bigger than the oven in the kitchen?", where the agent has to navigate to multiple locations ("dresser in bedroom", "oven in kitchen") and perform comparative reasoning ("dresser" bigger than "oven") before it can answer a question. Such questions require the development of entirely new modules or components in the agent. To address this, we propose a modular architecture composed of a program generator, a controller, a navigator, and a VQA module. The program generator converts the given question into sequential executable sub-programs; the navigator guides the agent to multiple locations pertinent to the navigation-related sub-programs; and the controller learns to select relevant observations along its path. These observations are then fed to the VQA module to predict the answer. We perform detailed analysis for each of the model components and show that our joint model can outperform previous methods and strong baselines by a significant margin.

Jianwei Yang, Zhile Ren, Mingze Xu et al.

Passive visual systems typically fail to recognize objects in the amodal setting where they are heavily occluded. In contrast, humans and other embodied agents have the ability to move in the environment, and actively control the viewing angle to better understand object shapes and semantics. In this work, we introduce the task of Embodied Visual Recognition (EVR): An agent is instantiated in a 3D environment close to an occluded target object, and is free to move in the environment to perform object classification, amodal object localization, and amodal object segmentation. To address this, we develop a new model called Embodied Mask R-CNN, for agents to learn to move strategically to improve their visual recognition abilities. We conduct experiments using the House3D environment. Experimental results show that: 1) agents with embodiment (movement) achieve better visual recognition performance than passive ones; 2) in order to improve visual recognition abilities, agents can learn strategical moving paths that are different from shortest paths.

Xinlei Chen, Ross Girshick, Kaiming He et al.

Sliding-window object detectors that generate bounding-box object predictions over a dense, regular grid have advanced rapidly and proven popular. In contrast, modern instance segmentation approaches are dominated by methods that first detect object bounding boxes, and then crop and segment these regions, as popularized by Mask R-CNN. In this work, we investigate the paradigm of dense sliding-window instance segmentation, which is surprisingly under-explored. Our core observation is that this task is fundamentally different than other dense prediction tasks such as semantic segmentation or bounding-box object detection, as the output at every spatial location is itself a geometric structure with its own spatial dimensions. To formalize this, we treat dense instance segmentation as a prediction task over 4D tensors and present a general framework called TensorMask that explicitly captures this geometry and enables novel operators on 4D tensors. We demonstrate that the tensor view leads to large gains over baselines that ignore this structure, and leads to results comparable to Mask R-CNN. These promising results suggest that TensorMask can serve as a foundation for novel advances in dense mask prediction and a more complete understanding of the task. Code will be made available.

Meet Shah, Xinlei Chen, Marcus Rohrbach et al.

Despite significant progress in Visual Question Answering over the years, robustness of today's VQA models leave much to be desired. We introduce a new evaluation protocol and associated dataset (VQA-Rephrasings) and show that state-of-the-art VQA models are notoriously brittle to linguistic variations in questions. VQA-Rephrasings contains 3 human-provided rephrasings for 40k questions spanning 40k images from the VQA v2.0 validation dataset. As a step towards improving robustness of VQA models, we propose a model-agnostic framework that exploits cycle consistency. Specifically, we train a model to not only answer a question, but also generate a question conditioned on the answer, such that the answer predicted for the generated question is the same as the ground truth answer to the original question. Without the use of additional annotations, we show that our approach is significantly more robust to linguistic variations than state-of-the-art VQA models, when evaluated on the VQA-Rephrasings dataset. In addition, our approach outperforms state-of-the-art approaches on the standard VQA and Visual Question Generation tasks on the challenging VQA v2.0 dataset.

Jin-Hwa Kim, Nikita Kitaev, Xinlei Chen et al.

In this work, we propose a goal-driven collaborative task that combines language, perception, and action. Specifically, we develop a Collaborative image-Drawing game between two agents, called CoDraw. Our game is grounded in a virtual world that contains movable clip art objects. The game involves two players: a Teller and a Drawer. The Teller sees an abstract scene containing multiple clip art pieces in a semantically meaningful configuration, while the Drawer tries to reconstruct the scene on an empty canvas using available clip art pieces. The two players communicate with each other using natural language. We collect the CoDraw dataset of ~10K dialogs consisting of ~138K messages exchanged between human players. We define protocols and metrics to evaluate learned agents in this testbed, highlighting the need for a novel "crosstalk" evaluation condition which pairs agents trained independently on disjoint subsets of the training data. We present models for our task and benchmark them using both fully automated evaluation and by having them play the game live with humans.

Gunnar A. Sigurdsson, Xinlei Chen, Abhinav Gupta

What does a typical visit to Paris look like? Do people first take photos of the Louvre and then the Eiffel Tower? Can we visually model a temporal event like "Paris Vacation" using current frameworks? In this paper, we explore how we can automatically learn the temporal aspects, or storylines of visual concepts from web data. Previous attempts focus on consecutive image-to-image transitions and are unsuccessful at recovering the long-term underlying story. Our novel Skipping Recurrent Neural Network (S-RNN) model does not attempt to predict each and every data point in the sequence, like classic RNNs. Rather, S-RNN uses a framework that skips through the images in the photo stream to explore the space of all ordered subsets of the albums via an efficient sampling procedure. This approach reduces the negative impact of strong short-term correlations, and recovers the latent story more accurately. We show how our learned storylines can be used to analyze, predict, and summarize photo albums from Flickr. Our experimental results provide strong qualitative and quantitative evidence that S-RNN is significantly better than other candidate methods such as LSTMs on learning long-term correlations and recovering latent storylines. Moreover, we show how storylines can help machines better understand and summarize photo streams by inferring a brief personalized story of each individual album.

Jiwei Li, Xinlei Chen, Eduard Hovy et al.

While neural networks have been successfully applied to many NLP tasks the resulting vector-based models are very difficult to interpret. For example it's not clear how they achieve {\em compositionality}, building sentence meaning from the meanings of words and phrases. In this paper we describe four strategies for visualizing compositionality in neural models for NLP, inspired by similar work in computer vision. We first plot unit values to visualize compositionality of negation, intensification, and concessive clauses, allow us to see well-known markedness asymmetries in negation. We then introduce three simple and straightforward methods for visualizing a unit's {\em salience}, the amount it contributes to the final composed meaning: (1) gradient back-propagation, (2) the variance of a token from the average word node, (3) LSTM-style gates that measure information flow. We test our methods on sentiment using simple recurrent nets and LSTMs. Our general-purpose methods may have wide applications for understanding compositionality and other semantic properties of deep networks , and also shed light on why LSTMs outperform simple recurrent nets,

Xinlei Chen, Abhinav Gupta

We present an approach to utilize large amounts of web data for learning CNNs. Specifically inspired by curriculum learning, we present a two-step approach for CNN training. First, we use easy images to train an initial visual representation. We then use this initial CNN and adapt it to harder, more realistic images by leveraging the structure of data and categories. We demonstrate that our two-stage CNN outperforms a fine-tuned CNN trained on ImageNet on Pascal VOC 2012. We also demonstrate the strength of webly supervised learning by localizing objects in web images and training a R-CNN style detector. It achieves the best performance on VOC 2007 where no VOC training data is used. Finally, we show our approach is quite robust to noise and performs comparably even when we use image search results from March 2013 (pre-CNN image search era).