LiDAR-based fully sparse architecture has garnered increasing attention. FSDv1 stands out as a representative work, achieving impressive efficacy and efficiency, albeit with intricate structures and handcrafted designs. In this paper, we present FSDv2, an evolution that aims to simplify the previous FSDv1 while eliminating the inductive bias introduced by its handcrafted instance-level representation, thus promoting better general applicability. To this end, we introduce the concept of \textbf{virtual voxels}, which takes over the clustering-based instance segmentation in FSDv1. Virtual voxels not only address the notorious issue of the Center Feature Missing problem in fully sparse detectors but also endow the framework with a more elegant and streamlined approach. Consequently, we develop a suite of components to complement the virtual voxel concept, including a virtual voxel encoder, a virtual voxel mixer, and a virtual voxel assignment strategy. Through empirical validation, we demonstrate that the virtual voxel mechanism is functionally similar to the handcrafted clustering in FSDv1 while being more general. We conduct experiments on three large-scale datasets: Waymo Open Dataset, Argoverse 2 dataset, and nuScenes dataset. Our results showcase state-of-the-art performance on all three datasets, highlighting the superiority of FSDv2 in long-range scenarios and its general applicability to achieve competitive performance across diverse scenarios. Moreover, we provide comprehensive experimental analysis to elucidate the workings of FSDv2. To foster reproducibility and further research, we have open-sourced FSDv2 at https://github.com/tusen-ai/SST.
Sound source localization aims to localize objects emitting the sound in visual scenes. Recent works obtaining impressive results typically rely on contrastive learning. However, the common practice of randomly sampling negatives in prior arts can lead to the false negative issue, where the sounds semantically similar to visual instance are sampled as negatives and incorrectly pushed away from the visual anchor/query. As a result, this misalignment of audio and visual features could yield inferior performance. To address this issue, we propose a novel audio-visual learning framework which is instantiated with two individual learning schemes: self-supervised predictive learning (SSPL) and semantic-aware contrastive learning (SACL). SSPL explores image-audio positive pairs alone to discover semantically coherent similarities between audio and visual features, while a predictive coding module for feature alignment is introduced to facilitate the positive-only learning. In this regard SSPL acts as a negative-free method to eliminate false negatives. By contrast, SACL is designed to compact visual features and remove false negatives, providing reliable visual anchor and audio negatives for contrast. Different from SSPL, SACL releases the potential of audio-visual contrastive learning, offering an effective alternative to achieve the same goal. Comprehensive experiments demonstrate the superiority of our approach over the state-of-the-arts. Furthermore, we highlight the versatility of the learned representation by extending the approach to audio-visual event classification and object detection tasks. Code and models are available at: https://github.com/zjsong/SACL.
Convolutional Neural Networks (CNNs) have been very successful at solving a variety of computer vision tasks such as object classification and detection, semantic segmentation, activity understanding, to name just a few. One key enabling factor for their great performance has been the ability to train very deep networks. Despite their huge success in many tasks, CNNs do not work well with non-Euclidean data, which is prevalent in many real-world applications. Graph Convolutional Networks (GCNs) offer an alternative that allows for non-Eucledian data input to a neural network. While GCNs already achieve encouraging results, they are currently limited to architectures with a relatively small number of layers, primarily due to vanishing gradients during training. This work transfers concepts such as residual/dense connections and dilated convolutions from CNNs to GCNs in order to successfully train very deep GCNs. We show the benefit of using deep GCNs (with as many as 112 layers) experimentally across various datasets and tasks. Specifically, we achieve very promising performance in part segmentation and semantic segmentation on point clouds and in node classification of protein functions across biological protein-protein interaction (PPI) graphs. We believe that the insights in this work will open avenues for future research on GCNs and their application to further tasks not explored in this paper. The source code for this work is available at https://github.com/lightaime/deep_gcns_torch and https://github.com/lightaime/deep_gcns for PyTorch and TensorFlow implementation respectively.
Computer vision has achieved great success in interpreting semantic meanings from images, yet estimating underlying (non-visual) physical properties of an object is often limited to their bulk values rather than reconstructing a dense map. In this work, we present our pressure eye (PEye) approach to estimate contact pressure between a human body and the surface she is lying on with high resolution from vision signals directly. PEye approach could ultimately enable the prediction and early detection of pressure ulcers in bed-bound patients, that currently depends on the use of expensive pressure mats. Our PEye network is configured in a dual encoding shared decoding form to fuse visual cues and some relevant physical parameters in order to reconstruct high resolution pressure maps (PMs). We also present a pixel-wise resampling approach based on Naive Bayes assumption to further enhance the PM regression performance. A percentage of correct sensing (PCS) tailored for sensing estimation accuracy evaluation is also proposed which provides another perspective for performance evaluation under varying error tolerances. We tested our approach via a series of extensive experiments using multimodal sensing technologies to collect data from 102 subjects while lying on a bed. The individual's high resolution contact pressure data could be estimated from their RGB or long wavelength infrared (LWIR) images with 91.8% and 91.2% estimation accuracies in $PCS_{efs0.1}$ criteria, superior to state-of-the-art methods in the related image regression/translation tasks.
N:M sparsity has received increasing attention due to its remarkable performance and latency trade-off compared with structured and unstructured sparsity. However, existing N:M sparsity methods do not differentiate the relative importance of weights among blocks and leave important weights underappreciated. Besides, they directly apply N:M sparsity to the whole network, which will cause severe information loss. Thus, they are still sub-optimal. In this paper, we propose an efficient and effective Multi-Axis Query methodology, dubbed as MaxQ, to rectify these problems. During the training, MaxQ employs a dynamic approach to generate soft N:M masks, considering the weight importance across multiple axes. This method enhances the weights with more importance and ensures more effective updates. Meanwhile, a sparsity strategy that gradually increases the percentage of N:M weight blocks is applied, which allows the network to heal from the pruning-induced damage progressively. During the runtime, the N:M soft masks can be precomputed as constants and folded into weights without causing any distortion to the sparse pattern and incurring additional computational overhead. Comprehensive experiments demonstrate that MaxQ achieves consistent improvements across diverse CNN architectures in various computer vision tasks, including image classification, object detection and instance segmentation. For ResNet50 with 1:16 sparse pattern, MaxQ can achieve 74.6\% top-1 accuracy on ImageNet and improve by over 2.8\% over the state-of-the-art. Codes and checkpoints are available at \url{https://github.com/JingyangXiang/MaxQ}.
The ambition of brain-inspired Spiking Neural Networks (SNNs) is to become a low-power alternative to traditional Artificial Neural Networks (ANNs). This work addresses two major challenges in realizing this vision: the performance gap between SNNs and ANNs, and the high training costs of SNNs. We identify intrinsic flaws in spiking neurons caused by binary firing mechanisms and propose a Spike Firing Approximation (SFA) method using integer training and spike-driven inference. This optimizes the spike firing pattern of spiking neurons, enhancing efficient training, reducing power consumption, improving performance, enabling easier scaling, and better utilizing neuromorphic chips. We also develop an efficient spike-driven Transformer architecture and a spike-masked autoencoder to prevent performance degradation during SNN scaling. On ImageNet-1k, we achieve state-of-the-art top-1 accuracy of 78.5\%, 79.8\%, 84.0\%, and 86.2\% with models containing 10M, 19M, 83M, and 173M parameters, respectively. For instance, the 10M model outperforms the best existing SNN by 7.2\% on ImageNet, with training time acceleration and inference energy efficiency improved by 4.5$\times$ and 3.9$\times$, respectively. We validate the effectiveness and efficiency of the proposed method across various tasks, including object detection, semantic segmentation, and neuromorphic vision tasks. This work enables SNNs to match ANN performance while maintaining the low-power advantage, marking a significant step towards SNNs as a general visual backbone. Code is available at https://github.com/BICLab/Spike-Driven-Transformer-V3.
4-dimensional (4D) radar is increasingly adopted in autonomous driving for perception tasks, owing to its robustness under adverse weather conditions. To better utilize the spatial information inherent in 4D radar data, recent deep learning methods have transitioned from using sparse point cloud to 4D radar tensors. However, the scarcity of publicly available 4D radar tensor datasets limits model generalization across diverse driving scenarios. Previous methods addressed this by synthesizing radar data, but the outputs did not fully exploit the spatial information characteristic of 4D radar. To overcome these limitations, we propose LiDAR-to-4D radar data synthesis (L2RDaS), a framework that synthesizes spatially informative 4D radar tensors from LiDAR data available in existing autonomous driving datasets. L2RDaS integrates a modified U-Net architecture to effectively capture spatial information and an object information supplement (OBIS) module to enhance reflection fidelity. This framework enables the synthesis of radar tensors across diverse driving scenarios without additional sensor deployment or data collection. L2RDaS improves model generalization by expanding real datasets with synthetic radar tensors, achieving an average increase of 4.25\% in ${{AP}_{BEV}}$ and 2.87\% in ${{AP}_{3D}}$ across three detection models. Additionally, L2RDaS supports ground-truth augmentation (GT-Aug) by embedding annotated objects into LiDAR data and synthesizing them into radar tensors, resulting in further average increases of 3.75\% in ${{AP}_{BEV}}$ and 4.03\% in ${{AP}_{3D}}$. The implementation will be available at https://github.com/kaist-avelab/K-Radar.
Vision Transformers (ViTs) are built on the assumption of treating image patches as ``visual tokens" and learn patch-to-patch attention. The patch embedding based tokenizer has a semantic gap with respect to its counterpart, the textual tokenizer. The patch-to-patch attention suffers from the quadratic complexity issue, and also makes it non-trivial to explain learned ViTs. To address these issues in ViT, this paper proposes to learn Patch-to-Cluster attention (PaCa) in ViT. Queries in our PaCa-ViT starts with patches, while keys and values are directly based on clustering (with a predefined small number of clusters). The clusters are learned end-to-end, leading to better tokenizers and inducing joint clustering-for-attention and attention-for-clustering for better and interpretable models. The quadratic complexity is relaxed to linear complexity. The proposed PaCa module is used in designing efficient and interpretable ViT backbones and semantic segmentation head networks. In experiments, the proposed methods are tested on ImageNet-1k image classification, MS-COCO object detection and instance segmentation and MIT-ADE20k semantic segmentation. Compared with the prior art, it obtains better performance in all the three benchmarks than the SWin and the PVTs by significant margins in ImageNet-1k and MIT-ADE20k. It is also significantly more efficient than PVT models in MS-COCO and MIT-ADE20k due to the linear complexity. The learned clusters are semantically meaningful. Code and model checkpoints are available at https://github.com/iVMCL/PaCaViT.
A fundamental and challenging problem in deep learning is catastrophic forgetting, i.e. the tendency of neural networks to fail to preserve the knowledge acquired from old tasks when learning new tasks. This problem has been widely investigated in the research community and several Incremental Learning (IL) approaches have been proposed in the past years. While earlier works in computer vision have mostly focused on image classification and object detection, more recently some IL approaches for semantic segmentation have been introduced. These previous works showed that, despite its simplicity, knowledge distillation can be effectively employed to alleviate catastrophic forgetting. In this paper, we follow this research direction and, inspired by recent literature on contrastive learning, we propose a novel distillation framework, Uncertainty-aware Contrastive Distillation (\method). In a nutshell, \method~is operated by introducing a novel distillation loss that takes into account all the images in a mini-batch, enforcing similarity between features associated to all the pixels from the same classes, and pulling apart those corresponding to pixels from different classes. In order to mitigate catastrophic forgetting, we contrast features of the new model with features extracted by a frozen model learned at the previous incremental step. Our experimental results demonstrate the advantage of the proposed distillation technique, which can be used in synergy with previous IL approaches, and leads to state-of-art performance on three commonly adopted benchmarks for incremental semantic segmentation. The code is available at \url{https://github.com/ygjwd12345/UCD}.
Pre-training & fine-tuning can enhance the transferring efficiency and performance in visual tasks. Recent delta-tuning methods provide more options for visual classification tasks. Despite their success, existing visual delta-tuning art fails to exceed the upper limit of full fine-tuning on challenging tasks like instance segmentation and semantic segmentation. To find a competitive alternative to full fine-tuning, we propose the Multi-cognitive Visual Adapter (Mona) tuning, a novel adapter-based tuning method. First, we introduce multiple vision-friendly filters into the adapter to enhance its ability to process visual signals, while previous methods mainly rely on language-friendly linear filters. Second, we add the scaled normalization layer in the adapter to regulate the distribution of input features for visual filters. To fully demonstrate the practicality and generality of Mona, we conduct experiments on multiple representative visual tasks, including instance segmentation on COCO, semantic segmentation on ADE20K, object detection on Pascal VOC, and image classification on several common datasets. Exciting results illustrate that Mona surpasses full fine-tuning on all these tasks and is the only delta-tuning method outperforming full fine-tuning on instance segmentation and semantic segmentation tasks. For example, Mona achieves a 1% performance gain on the COCO dataset compared to full fine-tuning. Comprehensive results suggest that Mona-tuning is more suitable for retaining and utilizing the capabilities of pre-trained models than full fine-tuning. The code will be released at https://github.com/Leiyi-Hu/mona.
Dense depth recovery is crucial in autonomous driving, serving as a foundational element for obstacle avoidance, 3D object detection, and local path planning. Adverse weather conditions, including haze, dust, rain, snow, and darkness, introduce significant challenges to accurate dense depth estimation, thereby posing substantial safety risks in autonomous driving. These challenges are particularly pronounced for traditional depth estimation methods that rely on short electromagnetic wave sensors, such as visible spectrum cameras and near-infrared LiDAR, due to their susceptibility to diffraction noise and occlusion in such environments. To fundamentally overcome this issue, we present a novel approach for robust metric depth estimation by fusing a millimeter-wave Radar and a monocular infrared thermal camera, which are capable of penetrating atmospheric particles and unaffected by lighting conditions. Our proposed Radar-Infrared fusion method achieves highly accurate and finely detailed dense depth estimation through three stages, including monocular depth prediction with global scale alignment, quasi-dense Radar augmentation by learning Radar-pixels correspondences, and local scale refinement of dense depth using a scale map learner. Our method achieves exceptional visual quality and accurate metric estimation by addressing the challenges of ambiguity and misalignment that arise from directly fusing multi-modal long-wave features. We evaluate the performance of our approach on the NTU4DRadLM dataset and our self-collected challenging ZJU-Multispectrum dataset. Especially noteworthy is the unprecedented robustness demonstrated by our proposed method in smoky scenarios. Our code will be released at \url{https://github.com/MMOCKING/RIDERS}.
Neural networks are fundamental in artificial intelligence, driving progress in computer vision and natural language processing. High-quality datasets are crucial for their development, and there is growing interest in datasets composed of neural networks themselves to support benchmarking, automated machine learning (AutoML), and model analysis. We introduce LEMUR, an open source dataset of neural network models with well-structured code for diverse architectures across tasks such as object detection, image classification, segmentation, and natural language processing. LEMUR is primarily designed to provide a rich source of structured model representations and associated performance data, enabling the fine-tuning of large language models for AutoML applications. Leveraging Python and PyTorch, LEMUR enables seamless extension to new datasets and models while maintaining consistency. It integrates an Optuna-powered framework for evaluation, hyperparameter optimization, statistical analysis, and graphical insights. LEMUR VR extension enables the seamless deployment of models in virtual reality, optimizing their performance on resource-constrained devices. Providing tools for model evaluation, preprocessing, and database management, LEMUR supports researchers and practitioners in developing, testing, and analyzing neural networks. It offers an API that delivers comprehensive information about neural network models and their complete performance statistics with a single request, which can be used in experiments with code-generating large language models. The LEMUR and its plugins are accessible as open source projects under the MIT license at https://github.com/ABrain-One/nn-dataset, https://github.com/ABrain-One/nn-plots and https://github.com/ABrain-One/nn-vr.
The recent increase in volume and complexity of available astronomical data has led to a wide use of supervised machine learning techniques. Active learning strategies have been proposed as an alternative to optimize the distribution of scarce labeling resources. However, due to the specific conditions in which labels can be acquired, fundamental assumptions, such as sample representativeness and labeling cost stability cannot be fulfilled. The Recommendation System for Spectroscopic follow-up (RESSPECT) project aims to enable the construction of optimized training samples for the Rubin Observatory Legacy Survey of Space and Time (LSST), taking into account a realistic description of the astronomical data environment. In this work, we test the robustness of active learning techniques in a realistic simulated astronomical data scenario. Our experiment takes into account the evolution of training and pool samples, different costs per object, and two different sources of budget. Results show that traditional active learning strategies significantly outperform random sampling. Nevertheless, more complex batch strategies are not able to significantly overcome simple uncertainty sampling techniques. Our findings illustrate three important points: 1) active learning strategies are a powerful tool to optimize the label-acquisition task in astronomy, 2) for upcoming large surveys like LSST, such techniques allow us to tailor the construction of the training sample for the first day of the survey, and 3) the peculiar data environment related to the detection of astronomical transients is a fertile ground that calls for the development of tailored machine learning algorithms.
To promote better performance-bandwidth trade-off for multi-agent perception, we propose a novel distilled collaboration graph (DiscoGraph) to model trainable, pose-aware, and adaptive collaboration among agents. Our key novelties lie in two aspects. First, we propose a teacher-student framework to train DiscoGraph via knowledge distillation. The teacher model employs an early collaboration with holistic-view inputs; the student model is based on intermediate collaboration with single-view inputs. Our framework trains DiscoGraph by constraining post-collaboration feature maps in the student model to match the correspondences in the teacher model. Second, we propose a matrix-valued edge weight in DiscoGraph. In such a matrix, each element reflects the inter-agent attention at a specific spatial region, allowing an agent to adaptively highlight the informative regions. During inference, we only need to use the student model named as the distilled collaboration network (DiscoNet). Attributed to the teacher-student framework, multiple agents with the shared DiscoNet could collaboratively approach the performance of a hypothetical teacher model with a holistic view. Our approach is validated on V2X-Sim 1.0, a large-scale multi-agent perception dataset that we synthesized using CARLA and SUMO co-simulation. Our quantitative and qualitative experiments in multi-agent 3D object detection show that DiscoNet could not only achieve a better performance-bandwidth trade-off than the state-of-the-art collaborative perception methods, but also bring more straightforward design rationale. Our code is available on https://github.com/ai4ce/DiscoNet.
Feature extraction and matching are the basic parts of many robotic vision tasks, such as 2D or 3D object detection, recognition, and registration. As is known, 2D feature extraction and matching have already achieved great success. Unfortunately, in the field of 3D, the current methods may fail to support the extensive application of 3D LiDAR sensors in robotic vision tasks due to their poor descriptiveness and inefficiency. To address this limitation, we propose a novel 3D feature representation method: Linear Keypoints representation for 3D LiDAR point cloud, called LinK3D. The novelty of LinK3D lies in that it fully considers the characteristics (such as the sparsity and complexity) of LiDAR point clouds and represents the keypoint with its robust neighbor keypoints, which provide strong constraints in the description of the keypoint. The proposed LinK3D has been evaluated on three public datasets, and the experimental results show that our method achieves great matching performance. More importantly, LinK3D also shows excellent real-time performance, faster than the sensor frame rate at 10 Hz of a typical rotating LiDAR sensor. LinK3D only takes an average of 30 milliseconds to extract features from the point cloud collected by a 64-beam LiDAR and takes merely about 20 milliseconds to match two LiDAR scans when executed on a computer with an Intel Core i7 processor. Moreover, our method can be extended to LiDAR odometry task, and shows good scalability. We release the implementation of our method at https://github.com/YungeCui/LinK3D.
Scene graph generation (SGG) of surgical procedures is crucial in enhancing holistically cognitive intelligence in the operating room (OR). However, previous works have primarily relied on multi-stage learning, where the generated semantic scene graphs depend on intermediate processes with pose estimation and object detection. This pipeline may potentially compromise the flexibility of learning multimodal representations, consequently constraining the overall effectiveness. In this study, we introduce a novel single-stage bi-modal transformer framework for SGG in the OR, termed S^2Former-OR, aimed to complementally leverage multi-view 2D scenes and 3D point clouds for SGG in an end-to-end manner. Concretely, our model embraces a View-Sync Transfusion scheme to encourage multi-view visual information interaction. Concurrently, a Geometry-Visual Cohesion operation is designed to integrate the synergic 2D semantic features into 3D point cloud features. Moreover, based on the augmented feature, we propose a novel relation-sensitive transformer decoder that embeds dynamic entity-pair queries and relational trait priors, which enables the direct prediction of entity-pair relations for graph generation without intermediate steps. Extensive experiments have validated the superior SGG performance and lower computational cost of S^2Former-OR on 4D-OR benchmark, compared with current OR-SGG methods, e.g., 3 percentage points Precision increase and 24.2M reduction in model parameters. We further compared our method with generic single-stage SGG methods with broader metrics for a comprehensive evaluation, with consistently better performance achieved.
There has been a significant increase in the adoption of technology in cricket recently. This trend has created the problem of duplicate work being done in similar computer vision-based research works. Our research tries to solve one of these problems by segmenting ball deliveries in a cricket broadcast using deep learning models, MobileNet and YOLO, thus enabling researchers to use our work as a dataset for their research. The output from our research can be used by cricket coaches and players to analyze ball deliveries which are played during the match. This paper presents an approach to segment and extract video shots in which only the ball is being delivered. The video shots are a series of continuous frames that make up the whole scene of the video. Object detection models are applied to reach a high level of accuracy in terms of correctly extracting video shots. The proof of concept for building large datasets of video shots for ball deliveries is proposed which paves the way for further processing on those shots for the extraction of semantics. Ball tracking in these video shots is also done using a separate RetinaNet model as a sample of the usefulness of the proposed dataset. The position on the cricket pitch where the ball lands is also extracted by tracking the ball along the y-axis. The video shot is then classified as a full-pitched, good-length or short-pitched delivery.
Neural architectures are the foundation for improving performance of deep neural networks (DNNs). This paper presents deep compositional grammatical architectures which harness the best of two worlds: grammar models and DNNs. The proposed architectures integrate compositionality and reconfigurability of the former and the capability of learning rich features of the latter in a principled way. We utilize AND-OR Grammar (AOG) as network generator in this paper and call the resulting networks AOGNets. An AOGNet consists of a number of stages each of which is composed of a number of AOG building blocks. An AOG building block splits its input feature map into N groups along feature channels and then treat it as a sentence of N words. It then jointly realizes a phrase structure grammar and a dependency grammar in bottom-up parsing the "sentence" for better feature exploration and reuse. It provides a unified framework for the best practices developed in state-of-the-art DNNs. In experiments, AOGNet is tested in the CIFAR-10, CIFAR-100 and ImageNet-1K classification benchmark and the MS-COCO object detection and segmentation benchmark. In CIFAR-10, CIFAR-100 and ImageNet-1K, AOGNet obtains better performance than ResNet and most of its variants, ResNeXt and its attention based variants such as SENet, DenseNet and DualPathNet. AOGNet also obtains the best model interpretability score using network dissection. AOGNet further shows better potential in adversarial defense. In MS-COCO, AOGNet obtains better performance than the ResNet and ResNeXt backbones in Mask R-CNN.
Image outpainting gains increasing attention since it can generate the complete scene from a partial view, providing a valuable solution to construct {360\textdegree} panoramic images. As image outpainting suffers from the intrinsic issue of unidirectional completion flow, previous methods convert the original problem into inpainting, which allows a bidirectional flow. However, we find that inpainting has its own limitations and is inferior to outpainting in certain situations. The question of how they may be combined for the best of both has as yet remained under-explored. In this paper, we provide a deep analysis of the differences between inpainting and outpainting, which essentially depends on how the source pixels contribute to the unknown regions under different spatial arrangements. Motivated by this analysis, we present a Cylin-Painting framework that involves meaningful collaborations between inpainting and outpainting and efficiently fuses the different arrangements, with a view to leveraging their complementary benefits on a seamless cylinder. Nevertheless, straightforwardly applying the cylinder-style convolution often generates visually unpleasing results as it discards important positional information. To address this issue, we further present a learnable positional embedding strategy to incorporate the missing component of positional encoding into the cylinder convolution, which significantly improves the panoramic results. It is noted that while developed for image outpainting, the proposed algorithm can be effectively extended to other panoramic vision tasks, such as object detection, depth estimation, and image super-resolution. Code will be made available at \url{https://github.com/KangLiao929/Cylin-Painting}.
Traditionally, convolutional neural networks (CNN) and vision transformers (ViT) have dominated computer vision. However, recently proposed vision graph neural networks (ViG) provide a new avenue for exploration. Unfortunately, for mobile applications, ViGs are computationally expensive due to the overhead of representing images as graph structures. In this work, we propose a new graph-based sparse attention mechanism, Sparse Vision Graph Attention (SVGA), that is designed for ViGs running on mobile devices. Additionally, we propose the first hybrid CNN-GNN architecture for vision tasks on mobile devices, MobileViG, which uses SVGA. Extensive experiments show that MobileViG beats existing ViG models and existing mobile CNN and ViT architectures in terms of accuracy and/or speed on image classification, object detection, and instance segmentation tasks. Our fastest model, MobileViG-Ti, achieves 75.7% top-1 accuracy on ImageNet-1K with 0.78 ms inference latency on iPhone 13 Mini NPU (compiled with CoreML), which is faster than MobileNetV2x1.4 (1.02 ms, 74.7% top-1) and MobileNetV2x1.0 (0.81 ms, 71.8% top-1). Our largest model, MobileViG-B obtains 82.6% top-1 accuracy with only 2.30 ms latency, which is faster and more accurate than the similarly sized EfficientFormer-L3 model (2.77 ms, 82.4%). Our work proves that well designed hybrid CNN-GNN architectures can be a new avenue of exploration for designing models that are extremely fast and accurate on mobile devices. Our code is publicly available at https://github.com/SLDGroup/MobileViG.