ACM MM 2024 Papers

The Vision Transformer has attained remarkable success in various computer vision applications. However, the large computational costs and complex design limit its ability in handling large feature maps. Existing research predominantly focuses on constraining attention to small local regions, which reduces the number of tokens attending the attention computation while overlooking computational demands caused by the feed-forward layer in the Vision Transformer block. In this paper, we introduce Group Vision Transformer (GVT), a relatively simple and efficient variant of Vision Transformer, aiming to improve attention computation. The core idea of our model is to divide and group the entire Transformer layer, instead of only the attention part, into multiple independent branches. This approach offers two advantages: (1) It helps reduce parameters and computational complexity; (2) it enhances the diversity of the learned features. We conduct comprehensive analysis of the impact of different numbers of groups on model performance, as well as their influence on parameters and computational complexity. Our proposed GVT demonstrates competitive performances in several common vision tasks. For example, our GVT-Tiny model achieves 84.8% top-1 accuracy on ImageNet-1K, 51.4% box mAP and 45.2% mask mAP on MS COCO object detection and instance segmentation, and 50.1% mIoU on ADE20K semantic segmentation, outperforming the CAFormer-S36 model by 0.3% in ImageNet-1K top-1 accuracy, 1.2% in box mAP, 1.0% in mask mAP on MS COCO object detection and instance segmentation, and 1.2% in mIoU on ADE20K semantic segmentation, with similar model parameters and computational complexity. Code is accessible at https://github.com/yaoppeng/GVT.

The audio-visual event localization task requires identifying concurrent visual and auditory events from unconstrained videos within a network model, locating them, and classifying their category. The efficient extraction and integration of audio and visual modal information have always been challenging in this field. In this paper, we introduce CACE-Net, which differs from most existing methods that solely use audio signals to guide visual information. We propose an audio-visual co-guidance attention mechanism that allows for adaptive bi-directional cross-modal attentional guidance between audio and visual information, thus reducing inconsistencies between modalities. Moreover, we have observed that existing methods have difficulty distinguishing between similar background and event and lack the fine-grained features for event classification. Consequently, we employ background-event contrast enhancement to increase the discrimination of fused feature and fine-tuned pre-trained model to extract more refined and discernible features from complex multimodal inputs. Specifically, we have enhanced the model's ability to discern subtle differences between event and background and improved the accuracy of event classification in our model. Experiments on the AVE dataset demonstrate that CACE-Net sets a new benchmark in the audio-visual event localization task, proving the effectiveness of our proposed methods in handling complex multimodal learning and event localization in unconstrained videos. Code is available at https://github.com/Brain-Cog-Lab/CACE-Net.

Recent advancements in 3D generation have garnered considerable interest due to their potential applications. Despite these advancements, the field faces persistent challenges in multi-conditional control, primarily due to the lack of paired datasets and the inherent complexity of 3D structures. To address these challenges, we introduce ImageBind3D, a novel framework for controllable 3D generation that integrates text, hand-drawn sketches, and depth maps to enhance user controllability. Our innovative contribution is adopting an inversion-align strategy, facilitating controllable 3D generation without requiring paired datasets. Firstly, utilizing GET3D as a baseline, our method innovates a 3D inversion technique that synchronizes 2D images with 3D shapes within the latent space of 3D GAN. Subsequently, we leverage images as intermediaries to facilitate pseudo-pairing between the shapes and various modalities. Moreover, our multi-modal diffusion model design strategically aligns external control signals with the generative model's latent knowledge, enabling precise and controllable 3D generation. Extensive experiments validate that ImageBind3D surpasses existing state-of-the-art methods in both fidelity and controllability. Additionally, our approach can offer composable guidance for any feed-forward 3D generative models, significantly enhancing their controllability.

Face anti-spoofing (FAS) based on domain generalization (DG) has attracted increasing attention from researchers.The reason for the poor generalization is that the model is overfitted to salient liveness-irrelevant signals.However, the previous methods alleviate the overfitting by mapping the images from multiple domains into a common feature space or promoting the separation of image features from domain-specific features and task-related features.If the text features of vision-language pre-trained (VLP) models (e.g., CLIP) are used to dynamically adjust the image features to gain a better generalization, we can not only explore a wider feature space but also avoid the potential degradation of semantic information.Specifically, we propose a FAS method of Style-Conditional Prompt Token Learning (S-CPTL), which aims to generate generalized text features by training the introduced prompt tokens to carry visual styles and use them as weights for classifiers to improve the model's generalization.Compared to the inherently static prompt token, we propose the dynamic prompt token, which can adaptively capture live-irrelevant signals from the instance-specific styles and increase their diversity through mixed feature statistics to further reduce the overfitting of the model.Thorough experimental analysis demonstrates that S-CPTL exceeds current top-performing methods in four distinct cross-dataset benchmarks.

Traditional deep learning models often struggle in few-shot learning scenarios, where limited labeled data is available. While the Contrastive Language-Image Pre-training (CLIP) model demonstrates impressive zero-shot capabilities, its performance in few-shot scenarios remains limited. Existing methods primarily aim to leverage the limited labeled dataset, but this offers limited potential for improvement. To overcome the limitations of small datasets in few-shot learning, we introduce a novel framework, SSAT-Adapter, that leverages CLIP's language understanding to generate informative auxiliary tasks and improve CLIP's performance and adaptability in few-shot settings. We utilize CLIP's language understanding to create decision-boundary-focused image latents. These latents form auxiliary tasks, including inter-class instances to bridge CLIP's pre-trained knowledge with the provided examples, and intra-class instances to subtly expand the representation of target classes. A self-paced training regime, progressing from easier to more complex tasks, further promotes robust learning. Experiments show our framework outperforms the state-of-the-art online few-shot learning method by an average of 2.2% on eleven image classification datasets. Further ablation studies on various tasks demonstrate the effectiveness of our approach to enhance CLIP's adaptability in few-shot image classification.

The perception of image aesthetics is built upon the understanding of semantic content. However, how to evaluate the aesthetic quality of images with diversified semantic backgrounds remains challenging in image aesthetics assessment (IAA). To address the dilemma, this paper presents a semantics-aware image aesthetics assessment approach, which first analyzes the semantic content of images and then models the aesthetic distinctions among images from two perspectives, i.e., aesthetic attribute and aesthetic level. Concretely, we propose two strategies, dubbed tag matching and contrastive ranking, to extract knowledge pertaining to image aesthetics. The tag matching identifies the semantic category and the dominant aesthetic attributes based on predefined tag libraries. The contrastive ranking is designed to uncover the comparative relationships among images with different aesthetic levels but similar semantic backgrounds. In the process of contrastive ranking, the impact of long-tailed distribution of aesthetic data is also considered by balanced sampling and traversal contrastive learning. Extensive experiments and comparisons on three benchmark IAA databases demonstrate the superior performance of the proposed model in terms of both prediction accuracy and alleviating long-tailed effect. The code will be public at https://github.com/yzc-ippl/TMCR **REMOVE 2nd URL**://github.com/yzc-ippl/TMCR.

Due to the limited permissions for upgrading dual-side (i.e., server-side and client-side) loss tolerance schemes from the perspective of CDN vendors in a multi-supplier market, modern large-scale live streaming services are still using the automatic-repeat-request (ARQ) based paradigm for loss recovery, which only requires server-side modifications. In this paper, we first conduct a large-scale measurement study with up to 50 million live streams. We find that loss shows dynamics and live streaming contains frequent on-off mode switching in the wild. We further find that the recovery latency, enlarged by the ubiquitous retransmission loss, is a critical factor affecting live streaming's client side QoE (e.g., video freezing). We then propose an enhanced recovery mechanism called AutoRec, which can transform the disadvantages of on-off mode switching into an advantage for reducing loss recovery latency without any modifications on the client side. AutoRec also adopts an online learning-based policy to fit the dynamics of loss, balancing the tradeoff between the recovery latency and the incurred overhead. We implement AutoRec upon QUIC and evaluate it via both testbed and real-world commercial services deployments. The experimental results demonstrate the practicability and profitability of AutoRec, in which the average times and duration of client-side video freezing can be lowered by 11.4% and 5.2%, respectively.

The development of person search techniques has been greatly promoted in recent years for its superior practicality and challenging goals. Despite their significant progress, existing person search models still lack the ability to continually learn from increasing real-world data and adaptively process input from different domains. To this end, this work introduces the continual person search task that sequentially learns on multiple domains and then performs person search on all seen domains. This requires balancing the stability and plasticity of the model to continually learn new knowledge without catastrophic forgetting. For this, we propose a Prompt-based Continual Person Search (PoPS) model in this paper. First, we design a compositional person search transformer to construct an effective pre-trained transformer without exhaustive pre-training from scratch on large-scale person search data. This serves as the fundamental for prompt-based continual learning. On top of that, we design a domain incremental prompt pool with a diverse attribute matching module. For each domain, we independently learn a set of prompts to encode the domain-oriented knowledge. Meanwhile, we jointly learn a group of diverse attribute projections and prototype embeddings to capture discriminative domain attributes. By matching an input image with the learned attributes across domains, the learned prompts can be properly selected for model inference. Extensive experiments are conducted to validate the proposed method for continual person search. The source code is available at https://github.com/PatrickZad/PoPS.

Accurate and robust multimodal multi-task perception is crucial for modern autonomous driving systems. However, current multimodal perception research follows independent paradigms designed for specific perception tasks, leading to a lack of complementary learning among tasks and decreased performance in multi-task learning (MTL) due to joint training. In this paper, we propose MaskBEV, a masked attention-based MTL paradigm that unifies 3D object detection and bird's eye view (BEV) map segmentation. MaskBEV introduces a task-agnostic Transformer decoder to process these diverse tasks, enabling MTL to be completed in a unified decoder without requiring additional design of specific task heads. To fully exploit the complementary information between BEV map segmentation and 3D object detection tasks in BEV space, we propose spatial modulation and scene-level context aggregation strategies. These strategies consider the inherent dependencies between BEV segmentation and 3D detection, naturally boosting MTL performance. Extensive experiments on nuScenes dataset show that compared with previous state-of-the-art MTL methods, MaskBEV achieves 1.3 NDS improvement in 3D object detection and 2.7 mIoU improvement in BEV map segmentation, while also demonstrating slightly leading inference speed.

Temporal relation modeling is one of the core aspects of few-shot action recognition. Most previous works mainly focus on temporal relation modeling based on coarse-level actions, without considering the atomic action details and fine-grained temporal information. This oversight represents a significant limitation in this task. Specifically, coarse-level temporal relation modeling can make the few-shot models overfit in high-discrepancy temporal context, and ignore the low-discrepancy but high-semantic relevance action details in the video. To address these issues, we propose a saliency-guided fine-grained temporal mask learning method that models the temporal atomic action relation for few-shot action recognition in a finer manner. First, to model the comprehensive temporal relations of video instances, we design a temporal mask learning architecture to automatically search for the best matching of each atomic action snippet. Next, to exploit the low-discrepancy atomic action features, we introduce a saliency-guided temporal mask module to adaptively locate and excavate the atomic action information. After that, the few-shot predictions can be obtained by feeding the embedded rich temporal-relation features to a common feature matcher. Extensive experimental results on standard datasets demonstrate our method's superior performance compared to existing state-of-the-art methods.

With the prosperity of the intelligent surveillance, multiple cameras have been applied to localize pedestrians more accurately. However, previous methods rely on laborious annotations of pedestrians in every frame and camera view. Therefore, we propose in this paper an Unsupervised Multi-view Pedestrian Detection approach (UMPD) to learn an annotation-free detector via vision-language models and 2D-3D cross-modal mapping: 1) Firstly, Semantic-aware Iterative Segmentation (SIS) is proposed to extract unsupervised representations of multi-view images, which are converted into 2D masks as pseudo labels, via our proposed iterative PCA and zero-shot semantic classes from vision-language models; 2) Secondly, we propose Geometry-aware Volume-based Detector (GVD) to end-to-end encode multi-view 2D images into a 3D volume to predict voxel-wise density and color via 2D-to-3D geometric projection, trained by 3D-to-2D rendering losses with SIS pseudo labels; 3) Thirdly, for better detection results, i.e., the 3D density projected on Birds-Eye-View, we propose Vertical-aware BEV Regularization (VBR) to constrain pedestrians to be vertical like the natural poses. Extensive experiments on popular multi-view pedestrian detection benchmarks Wildtrack, Terrace, and MultiviewX, show that our proposed UMPD, as the first fully-unsupervised method to our best knowledge, performs competitively to the previous state-of-the-art supervised methods. Code is available at https://github.com/lmy98129/UMPD.

With the advancement of AIGC, video frame interpolation (VFI) has become a crucial component in existing video generation frameworks, attracting widespread research interest. For the VFI task, the motion estimation between neighboring frames plays a crucial role in avoiding motion ambiguity. However, existing VFI methods always struggle to accurately predict the motion information between consecutive frames, and this imprecise estimation leads to blurred and visually incoherent interpolated frames. In this paper, we propose a novel diffusion framework, Motion-Aware latent Diffusion models (MADiff), which is specifically designed for the VFI task. By incorporating motion priors between the conditional neighboring frames with the target interpolated frame predicted throughout the diffusion sampling procedure, MADiff progressively refines the intermediate outcomes, culminating in generating both visually smooth and realistic results. Extensive experiments conducted on benchmark datasets demonstrate that our method achieves state-of-the-art performance significantly outperforming existing approaches, especially under challenging scenarios involving dynamic textures with complex motion.

Unsupervised Visible-Infrared Person Re-identification (USVI-ReID) presents a formidable challenge, which aims to match pedestrian images across visible and infrared modalities without any annotations. Recently, clustered pseudo-label methods have become predominant in USVI-ReID, although the inherent noise in pseudo-labels presents a significant obstacle. Most existing works primarily focus on shielding the model from the harmful effects of noise, neglecting to calibrate noisy pseudo-labels usually associated with hard samples, which will compromise the robustness of the model. To address this issue, we design a Robust Pseudo-label Learning with Neighbor Relation (RPNR) framework for USVI-ReID. To be specific, we first introduce a straightforward yet potent Noisy Pseudo-label Calibration module to correct noisy pseudo-labels. Due to the high intra-class variations, noisy pseudo-labels are difficult to calibrate completely. Therefore, we introduce a Neighbor Relation Learning module to reduce high intra-class variations by modeling potential interactions between all samples. Subsequently, we devise an Optimal Transport Prototype Matching module to establish reliable cross-modality correspondences. On that basis, we design a Memory Hybrid Learning module to jointly learn modality-specific and modality-invariant information. Comprehensive experiments conducted on two widely recognized benchmarks, SYSU-MM01 and RegDB, demonstrate that RPNR outperforms the current state-of-the-art GUR with an average Rank-1 improvement of 10.3%. The code is available at https://github.com/XiangboYin/RPNR.

3D Gaussian Splatting (3D-GS) technique couples 3D Gaussian primitives with differentiable rasterization to achieve high-quality novel view synthesis results while providing advanced real-time rendering performance. However, due to the flaw of its adaptive density control strategy in 3D-GS, it frequently suffers from over-reconstruction issue in intricate scenes containing high-frequency details, leading to blurry rendered images. The underlying reason for the flaw has still been under-explored. In this work, we present a comprehensive analysis of the cause of aforementioned artifacts, namely gradient collision, which prevents large Gaussians in over-reconstructed regions from splitting. To address this issue, we propose the novel homodirectional view-space positional gradient as the criterion for densification. Our strategy efficiently identifies large Gaussians in over-reconstructed regions, and recovers fine details by splitting. We evaluate our proposed method on various challenging datasets. The experimental results indicate that our approach achieves the best rendering quality with reduced or similar memory consumption. Our method is easy to implement and can be incorporated into a wide variety of most recent Gaussian Splatting-based methods. We will open source our codes upon formal publication. Our project page is available at: https://ty424.github.io/AbsGS.github.io/

Pixel-level fine-grained image editing remains an open challenge. Previous works fail to achieve an ideal trade-off between control granularity and inference speed. They either fail to achieve pixel-level fine-grained control, or their inference speed requires optimization. To address this, this paper for the first time employs a regression-based network to learn the variation patterns of StyleGAN latent codes during the image dragging process. This method enables pixel-level precision in dragging editing with little time cost. Users can specify handle points and their corresponding target points on any GAN-generated images, and our method will move each handle point to its corresponding target point. Through experimental analysis, we discover that a short movement distance from handle points to target points yields a high-fidelity edited image, as the model only needs to predict the movement of a small portion of pixels. To achieve this, we decompose the entire movement process into multiple sub-processes. Specifically, we develop a transformer encoder-decoder based network named 'Latent Predictor' to predict the latent code motion trajectories from handle points to target points in an autoregressive manner. Moreover, to enhance the prediction stability, we introduce a component named 'Latent Regularizer', aimed at constraining the latent code motion within the distribution of natural images. Extensive experiments demonstrate that our method achieves state-of-the-art (SOTA) inference speed and image editing performance at the pixel-level granularity.

Point cloud segmentation forms the foundation of 3D scene understanding. Boundaries, the intersections of regions, are prone to mis-segmentation. Current point cloud segmentation models exhibit unsatisfactory performance on boundaries. There is limited focus on explicitly addressing semantic segmentation of point cloud boundaries. We introduce a method called Multi-fineness Boundary Constraint (MBC) to tackle this challenge. By querying boundaries at various degrees of fineness and imposing feature constraints within these boundary areas, we enhance the discrimination between boundaries and non-boundaries, improving point cloud boundary segmentation. However, solely emphasizing boundaries may compromise the segmentation accuracy in broader non-boundary regions. To mitigate this, we introduce a new concept of point cloud space termed ensemble and a Shifted Ensemble-aware Perception (SEP) module. This module establishes information interactions between points with minimal computational cost, effectively capturing direct point-to-point long-range correlations within ensembles. It enhances segmentation performance for both boundaries and non-boundaries.

Large vision-language models (LVLMs) integrate visual information into large language models, showcasing remarkable multi-modal conversational capabilities. However, the visual modules introduces new challenges in terms of robustness for LVLMs, as attackers can craft adversarial images that are visually clean but may mislead the model to generate incorrect answers. In general, LVLMs rely on vision encoders to transform images into visual tokens, which are crucial for the language models to perceive image contents effectively. Therefore, we are curious about one question: Can LVLMs still generate correct responses when the encoded visual tokens are attacked and disrupting the visual information? To this end, we propose a non-targeted attack method referred to as VT-Attack (Visual Tokens Attack), which constructs adversarial examples from multiple perspectives, with the goal of comprehensively disrupting feature representations and inherent relationships as well as the semantic properties of visual tokens output by image encoders. Using only access to the image encoder in the proposed attack, the generated adversarial examples exhibit transferability across diverse LVLMs utilizing the same image encoder and generality across different tasks. Extensive experiments validate the superior attack performance of the VT-Attack over baseline methods, demonstrating its effectiveness in attacking LVLMs with image encoders, which in turn can provide guidance on the robustness of LVLMs, particularly in terms of the stability of the visual feature space.

Few-Shot Learning (FSL) aims to recognize new classes with limited labeled data. Recent studies have attempted to address the challenge of rare samples with textual prompts to modulate visual features. However, they usually struggle to capture complex semantic relationships between textual and visual features. Moreover, vanilla self-attention is heavily affected by useless information in images, severely constraining the potential of semantic priors in FSL due to the confusion of numerous irrelevant tokens during interaction. To address these aforementioned issues, a K-NN Transformer with Pyramid Prompts (KTPP) is proposed to select discriminative information with K-NN Context Attention (KCA) and adaptively modulate visual features with Pyramid Cross-modal Prompts (PCP). First, for each token, the KCA only selects the K most relevant tokens to compute the self-attention matrix and incorporates the mean of all tokens as the context prompt to provide the global context in three cascaded stages. As a result, irrelevant tokens can be progressively suppressed. Secondly, pyramid prompts are introduced in the PCP to emphasize visual features via interactions between text-based class-aware prompts and multi-scale visual features. This allows the ViT to dynamically adjust the importance weights of visual features based on rich semantic information at different scales, making models robust to spatial variations. Finally, augmented visual features and class-aware prompts are interacted via the KCA to extract class-specific features. Consequently, our model further enhances noise-free visual representations via deep cross-modal interactions, extracting generalized visual representation in scenarios with few labeled samples. Extensive experiments on four benchmark datasets demonstrate significant gains over the state-of-the-art methods, especially for the 1-shot task with 2.28% improvement on average due to semantically enhanced visual representations.