ACM MM 2024 Papers

The recent advances in query-based multi-camera 3D object detection are featured by initializing object queries in the 3D space, and then sampling features from perspective-view images to perform multi-round query refinement. In such a framework, query points near the same camera ray are likely to sample similar features from very close pixels, resulting in ambiguous query features and degraded detection accuracy. To this end, we introduce RayFormer, a camera-ray-inspired query-based 3D object detector that aligns the initialization and feature extraction of object queries with the optical characteristics of cameras. Specifically, RayFormer transforms perspective-view image features into bird's eye view (BEV) via the lift-splat-shoot method and segments the BEV map to sectors based on the camera rays. Object queries are uniformly and sparsely initialized along each camera ray, facilitating the projection of different queries onto different areas in the image to extract distinct features. Besides, we leverage the instance information of images to supplement the uniformly initialized object queries by further involving additional queries along the ray from 2D object detection boxes. To extract unique object-level features that cater to distinct queries, we design a ray sampling method that suitably organizes the distribution of feature sampling points on both images and bird's eye view. Extensive experiments are conducted on the nuScenes dataset to validate our proposed ray-inspired model design. The proposed RayFormer achieves 55.5% mAP and 63.3% NDS, respectively.

Co-Speech gesture generation encounters challenges with imbalanced, long-tailed gesture distributions. While recent methods typically address this by employing Vector Quantized Variational Autoencoder (VQ-VAE), encode gestures into a codebook and classify codebook indices based on audio or text cues. However, due to the imbalanced, the codebook classification tends to bias towards majority gestures, neglecting semantically rich minority gestures. To address this, this paper proposes the Entropy-Guided Co-Speech Gesture Generation (EGGesture). EGGesture leverages an Entropy-Guided VQ-VAE to jointly optimizes the distribution of codebook indices and adjusts loss weights for codebook index classification, which consists of a) A differentiable approach for entropy computation using Gumbel-Softmax and cosine similarity, facilitating online codebook distribution optimization, and b) a strategy that utilizes computed codebook entropy to collaboratively guide the classification loss weighting. These designs enable the dynamic refinement of the codebook utilization, striking a balance between the quality of the learned gesture representation and the accuracy of the classification phase. Experiments on the Trinity and BEAT datasets demonstrate EGGesture's state-of-the-art performance both qualitatively and quantitatively.

Cross-modal coherence modeling is essential for intelligent systems to help them organize and structure information, thereby understanding and creating content of the physical world coherently like human-beings. Previous work on cross-modal coherence modeling attempted to leverage the order information from another modality to assist the coherence recovering of the target modality. Despite of the effectiveness, labeled associated coherency information is not always available and might be costly to acquire, making the cross-modal guidance hard to leverage. To tackle this challenge, this paper explores a new way to take advantage of cross-modal guidance without gold labels on coherency, and proposes the Weak Cross-Modal Guided Ordering (WeGO) model. More specifically, it leverages high-confidence predicted pairwise order in one modality as reference information to guide the coherence modeling in another. An iterative learning paradigm is further designed to jointly optimize the coherence modeling in two modalities with selected guidance from each other. The iterative cross-modal boosting also functions in inference to further enhance coherence prediction in each modality. Experimental results on two public datasets have demonstrated that the proposed method outperforms existing methods for cross-modal coherence modeling tasks. Major technical modules have been evaluated effective through ablation studies. Codes are available at: \url{https://github.com/scvready123/IterWeGO}.

Exploring open-vocabulary video action recognition is a promising venture, which aims to recognize previously unseen actions within any arbitrary set of categories. Existing methods typically adapt pretrained image-text models to the video domain, capitalizing on their inherent strengths in generalization. A common thread among such methods is the augmentation of visual embeddings with temporal information to improve the recognition of seen actions. Yet, they compromise with standard less-informative action descriptions, thus faltering when confronted with novel actions. Drawing inspiration from human cognitive processes, we argue that augmenting text embeddings with human prior knowledge is pivotal for open-vocabulary video action recognition. To realize this, we innovatively blend video models with Large Language Models (LLMs) to devise Action-conditioned Prompts. Specifically, we harness the knowledge in LLMs to produce a set of descriptive sentences that contain distinctive features for identifying given actions. Building upon this foundation, we further introduce a multi-modal action knowledge alignment mechanism to align concepts in video and textual knowledge encapsulated within the prompts. Extensive experiments on various video benchmarks, including zero-shot, few-shot, and base-to-novel generalization settings, demonstrate that our method not only sets new SOTA performance but also possesses excellent interpretability.

Acquiring commonsense knowledge about entity-pairs from images is crucial across diverse applications. Distantly supervised learning has made significant advancements by automatically retrieving images containing entity pairs and summarizing commonsense knowledge from the bag of images. However, the retrieved images may not always cover all possible relations, and the informative features across the bag of images are often overlooked. To address these challenges, a Multi-modal Cross-domain Feature Learning framework is proposed to incorporate the general domain knowledge from a large vision-text foundation model, ViT-GPT2, to handle unseen relations and exploit complementary information from multiple sources. Then, a Group Attention module is designed to exploit the attentive information from other instances of the same bag to boost the informative features of individual instances. Finally, a Gamma-corrected Gated Fusion is designed to select a subset of informative instances for a comprehensive summarization of commonsense entity relations. Extensive experimental results demonstrate the superiority of the proposed method over state-of-the-art models for extracting commonsense knowledge.

Recently, transformers have demonstrated great potential for modeling long-term dependencies from skeleton sequences and thereby gained ever-increasing attention in skeleton action recognition. However, the existing transformer-based approaches heavily rely on the naive attention mechanism for capturing the spatiotemporal features, which falls short in learning discriminative representations that exhibit similar motion patterns. To address this challenge, we introduce the Frequency-aware Mixed Transformer (FreqMixFormer), specifically designed for recognizing similar skeletal actions with subtle discriminative motions. First, we introduce a frequency-aware attention module to unweave skeleton frequency representations by embedding joint features into frequency attention maps, aiming to distinguish the discriminative movements based on their frequency coefficients. Subsequently, we develop a mixed transformer architecture to incorporate spatial features with frequency features to model the comprehensive frequency-spatial patterns. Additionally, a temporal transformer is proposed to extract the global correlations across frames. Extensive experiments show that FreqMiXFormer outperforms SOTA on 3 popular skeleton action recognition datasets, including NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets. Our project is publicly available at: https://github.com/wenhanwu95/FreqMixFormer.

Pre-trained language models (PLMs) that rely solely on textual corpus may present limitations in multimodal semantics comprehension. Existing studies attempt to alleviate this issue by incorporating additional modal information through image retrieval or generation. However, these methods: (1) inevitably encounter modality gaps and noise; (2) treat all modalities indiscriminately; and (3) ignore visual or acoustic semantics of key entities. To tackle these challenges, we propose a novel principled iterative framework for multimodal-augmented PLMs termed MASE, which achieves efficient and balanced injection of multimodal semantics under the proposed Expectation Maximization (EM) based iterative algorithm. Initially, MASE utilizes multimodal proxies instead of explicit data to enhance PLMs, which avoids noise and modality gaps. In E-step, MASE adopts a novel information-driven self-balanced strategy to estimate allocation weights. Furthermore, MASE employs heterogeneous graph attention to capture entity-level fine-grained semantics on the proposed multimodal-semantic scene graph. In M-step, MASE injects global multimodal knowledge into PLMs through a cross-modal contrastive loss. Experimental results show that MASE consistently outperforms competitive baselines on multiple tasks across various architectures. More impressively, MASE is compatible with existing efficient parameter fine-tuning methods, such as prompt learning.

Recent advances in multimodal artificial intelligence have greatly improved the integration of vision-language-audio cues to enrich the content creation process. Inspired by these developments, in this paper, we first integrate audio into the face inpainting task to facilitate identity manipulation. Our main insight is that a person's voice carries distinct identity markers, such as age and gender, which provide an essential supplement for identity-aware face inpainting. By extracting identity information from audio as guidance, our method can naturally support tasks of identity preservation and identity swapping in face inpainting. Specifically, we introduce a dual-stream network architecture comprising a face branch and an audio branch. The face branch is tasked with extracting deterministic information from the visible parts of the input masked face, while the audio branch is designed to capture heuristic identity priors from the speaker's voice. The identity codes from two streams are integrated using a multi-layer perceptron (MLP) to create a virtual unified identity embedding that represennts comprehensive identity features. In addition, to explicitly exploit the information from audio, we introduce an audio-face generator to generate an 'fake' audio face directly from audio and fuse the multi-scale intermediate features from the audio-face generator into face inpainting network through an audio-visual feature fusion (AVFF) module. Extensive experiments demonstrate the positive impact of extracting identity information from audio on face inpainting task, especially in identity preservation.

Droplet-based microfluidic devices, with their high throughput and low power consumption, have found wide-ranging applications in the life sciences, such as drug discovery and cancer detection. However, the lack of real-time methods for accurately estimating droplet generation parameters has resulted in droplet microfluidic systems remaining largely offline-controlled, making it challenging to achieve efficient feedback in droplet generation. To meet the real-time requirements, it's imperative to minimize the data throughput of the collection system while employing parameter estimation algorithms that are both resource-efficient and highly effective. Spike camera, as an innovative form of neuromorphic camera, facilitates high temporal resolution scene capture with comparatively low data throughput. In this paper, we propose a real-time evaluation method for high-speed droplet parameters based on spike-based microfluidic flow-focusing, named RTDE, that integrates spike camera into the droplet collection system to efficiently capture information using spike stream. To process the spike stream effectively, we develop a spike-based estimation algorithm for real-time droplet generation parameters. To validate the performance of our method, we collected spike-based droplet datasets (SDD), comprising synthetic and real data with varying flow velocities, frequencies, and droplet sizes. Experiments result on these datasets consistently demonstrate that our method achieves parameter estimations that closely match the ground truth values, showcasing high precision. Furthermore, comparative experiments with image-based parameter estimation methods highlight the superior time efficiency of our method, enabling real-time calculation of parameter estimations. Cdoe and datasets are avaliable at: https://github.com/Onetism/RTDE

High-resolution point clouds (HRPCD) anomaly detection (AD) plays a critical role in precision machining and high-end equipment manufacturing. Despite considerable 3D-AD methods that have been proposed recently, they still cannot meet the requirements of the HRPCD-AD task. There are several challenges: i) It is difficult to directly capture HRPCD information due to large amounts of points at the sample level; ii) The advanced transformer-based methods usually obtain anisotropic features, leading to degradation of the representation; iii) The proportion of abnormal areas is very small, which makes it difficult to characterize. To address these challenges, we propose a novel group-level feature-based network, called Group3AD, which has a significantly efficient representation ability. First, we design an Intercluster Uniformity Network (IUN) to present the mapping of different groups in the feature space as several clusters, and obtain a more uniform distribution between clusters representing different parts of the point clouds in the feature space. Then, an Intracluster Alignment Network (IAN) is designed to encourage groups within the cluster to be distributed tightly in the feature space. In addition, we propose an Adaptive Group-Center Selection (AGCS) based on geometric information to improve the pixel density of potential anomalous regions during inference. The experimental results verify the effectiveness of our proposed Group3AD, which surpasses Reg3D-AD by the margin of 5% in terms of object-level AUROC on Real3D-AD. We provide the code and supplementary information on our website: https://github.com/M-3LAB/Group3AD.

Multimodal sentiment analysis (MSA) aims to predict sentiment from text, audio, and visual data of videos. Existing works focus on designing fusion strategies or decoupling mechanisms, which suffer from low data utilization and a heavy reliance on large amounts of labeled data. However, acquiring large-scale annotations for multimodal sentiment analysis is extremely labor-intensive and costly. To address this challenge, we propose GRACE, a GRadient-based Active learning method with Curriculum Enhancement, designed for MSA under a multi-task learning framework. Our approach achieves annotation reduction by strategically selecting valuable samples from the unlabeled data pool while maintaining high-performance levels. Specifically, we introduce informativeness and representativeness criteria, calculated from gradient magnitudes and sample distances, to quantify the active value of unlabeled samples. Additionally, an easiness criterion is incorporated to avoid outliers, considering the relationship between modality consistency and sample difficulty. During the learning process, we dynamically balance sample difficulty and active value, guided by the curriculum learning principle. This strategy prioritizes easier, modality-aligned samples for stable initial training, then gradually increases the difficulty by incorporating more challenging samples with modality conflicts. Extensive experiments demonstrate the effectiveness of our approach on both multimodal sentiment regression and classification benchmarks.

Insufficient labeled training samples pose a critical challenge in multi-label classification, potentially leading to overfitting of the model. This paper delineates a criterion for establishing a common domain among different datasets, whereby datasets sharing analogous object descriptions and label structures are considered part of the 'same field'. Integrating samples from disparate datasets within this shared field for training purposes effectively mitigates overfitting and enhances model accuracy. Motivated by this approach, we introduce a novel method for multi-label classification termed Non-Overlapped Multi-View Weak-Label Learning Guided by Multiple Correlations (NOMWM). Our method strategically amalgamates samples from diverse datasets within the shared field to enrich the training dataset. Furthermore, we project samples from various datasets onto a unified subspace to facilitate learning in a consistent latent space. Additionally, we address the challenge of weak labels stemming from incomplete label overlaps across datasets. Leveraging weak-label indicator matrices and label correlation mining techniques, we effectively mitigate the impact of weak labels. Extensive experimentation on multiple benchmark datasets validates the efficacy of our method, demonstrating clear improvements over existing state-of-the-art approaches.

Medical report generation aims at automating the synthesis of accurate and comprehensive diagnostic reports from radiological images. The task can significantly enhance clinical decision-making and alleviate the workload on radiologists. Existing works normally generate reports from single chest radiographs, although historical examination data also serve as crucial references for radiologists in real-world clinical settings. To address this constraint, we introduce a novel framework that mimics the workflow of radiologists. This framework compares past and present patient images to monitor disease progression and incorporates prior diagnostic reports as references for generating current personalized reports. We tackle the textual diversity challenge in cross-modal tasks by promoting style-agnostic discrete report representation learning and token generation. Furthermore, we propose a novel spatio-temporal fusion method with multi-granularities to fuse textual and visual features by disentangling the differences between current and historical data. We also tackle token generation biases, which arise from long-tail frequency distributions, proposing a novel feature normalization technique. This technique ensures unbiased generation for tokens, whether they are frequent or infrequent, enabling the robustness of report generation for rare diseases. Experimental results on the two public datasets demonstrate that our proposed model outperforms state-of-the-art baselines.

Recently, large-scale visual language pre-trained (VLP) models have demonstrated impressive performance across various downstream tasks. Motivated by these advancements, pioneering efforts have emerged in multi-label image recognition with missing labels, leveraging VLP prompt-tuning technology. However, they usually cannot match text and vision features well, due to complicated semantics gaps and missing labels in a multi-label image. To tackle this challenge, we propose Text-Region Matching for optimizing Multi-Label prompt tuning, namely TRM-ML, a novel method for enhancing meaningful cross-modal matching. Compared to existing methods, we advocate exploring the information of category-aware regions rather than the entire image or pixels, which contributes to bridging the semantic gap between textual and visual representations in a one-to-one matching manner. Concurrently, we further introduce multimodal contrastive learning to narrow the semantic gap between textual and visual modalities and establish intra-class and inter-class relationships. Additionally, to deal with missing labels, we propose a multimodal category prototype that leverages intra- and inter-category semantic relationships to estimate unknown labels, facilitating pseudo-label generation. Extensive experiments on the MS-COCO, PASCAL VOC, Visual Genome, NUS-WIDE, and CUB-200-211 benchmark datasets demonstrate that our proposed framework outperforms the state-of-the-art methods by a significant margin. Our code is available here.

Geometry and color information provided by the point clouds are both crucial for 3D scene understanding. Two pieces of information characterize the different aspects of point clouds, but existing methods lack an elaborate design for the discrimination and relevance. Hence we explore a 3D self-supervised paradigm that can better utilize the relations of point cloud information. Specifically, we propose a universal 3D scene pre-training framework via Geometry-Color Contrast (Point-GCC), which aligns geometry and color information using a Siamese network. To take care of actual application tasks, we design (i) hierarchical supervision with point-level contrast and reconstruct and object-level contrast based on the novel deep clustering module to close the gap between pre-training and downstream tasks; (ii) architecture-agnostic backbone to adapt for various downstream models. Benefiting from the object-level representation associated with downstream tasks, Point-GCC can directly evaluate model performance and the result demonstrates the effectiveness of our methods. Transfer learning results on a wide range of tasks also show consistent improvements across all datasets. e.g., new state-of-the-art object detection results on SUN RGB-D and S3DIS datasets. Codes will be released at https://github.com/Asterisci/Point-GCC.

Learning multiple proxy tasks is a popular training strategy in semi-supervised video anomaly detection. However, the traditional method of learning multiple proxy tasks simultaneously is prone to suboptimal solutions, and simply executing multiple proxy tasks sequentially cannot ensure continuous performance improvement. In this paper, we thoroughly investigate the impact of task composition and training order on performance enhancement. We find that ensuring continuous performance improvement in multi-task learning requires different but continuous optimization objectives in different training phases. To this end, a training strategy based on progressive learning is proposed to enhance the multi-task learning in VAD. The learning objectives of the model in previous phases contribute to the training in subsequent phases. Specifically, we decompose video anomaly detection into three phases: perception, comprehension, and inference, continuously refining the learning objectives to enhance model performance. In the three phases, we perform the visual task, the semantic task and the open-set task in turn to train the model. The model learns different levels of features and focuses on different types of anomalies in different phases. Extensive experiments demonstrate the effectiveness of our method, highlighting that the benefits derived from the progressive learning transcend specific proxy tasks.

Long-term time series forecasting is a long-standing challenge in various applications. A central issue in time series forecasting is that methods should expressively capture long-term dependency. Furthermore, time series forecasting methods should be flexible when applied to different scenarios. Although Fourier analysis offers an alternative to effectively capture reusable and periodic patterns to achieve long-term forecasting in different scenarios, existing methods often assume high-frequency components represent noise and should be discarded in time series forecasting. However, we conduct a series of motivation experiments and discover that the role of certain frequencies varies depending on the scenarios. In some scenarios, removing high-frequency components from the original time series can improve the forecasting performance, while in others scenarios, removing them is harmful to forecasting performance. Therefore, it is necessary to treat the frequencies differently according to specific scenarios. To achieve this, we first reformulate the time series forecasting problem as learning a transfer function of each frequency in the Fourier domain. Further, we design Frequency Dynamic Fusion (FreDF), which individually predicts each Fourier component, and dynamically fuses the output of different frequencies. Moreover, we provide a novel insight into the generalization ability of time series forecasting and propose the generalization bound of time series forecasting. Then we prove FreDF has a lower bound, indicating that FreDF has better generalization ability. Extensive experiments conducted on multiple benchmark datasets and ablation studies demonstrate the effectiveness of FreDF.

The burgeoning field of text-to-music generation models has shown great promise in their ability to generate high-quality music aligned with users' textual descriptions. These models effectively capture abstract/global musical features such as style and mood. However, they often inadequately produce the precise rendering of critical music loop attributes, including melody, rhythms, and instrumentation, which are essential for modern music loop production. To overcome this limitation, this paper proposed a Loops Transformer and a Multi-Stage Cross Attention mechanism that enable a cohesive integration of textual and MIDI input specifications. Additionally, a novel Instrument-Aware Reinforcement Learning technique was introduced to ensure the correct adoption of instrumentation. We demonstrated that the proposed model can generate music loops that simultaneously satisfy the conditions specified by both natural language texts and MIDI input, ensuring coherence between the two modalities. We also showed that our model outperformed the state-of-the-art baseline model, MusicGen, in both objective metrics (by lowering the FAD score by 1.3, indicating superior quality with lower scores, and by improving the Normalized Dynamic Time Warping Distance with given melodies by 12%) and subjective metrics (by +2.56% in OVL, +5.42% in REL, and +7.74% in Loop Consistency). These improvements highlight our model's capability to produce musically coherent loops that satisfy the complex requirements of contemporary music production, representing a notable advancement in the field. Generated music loop samples can be explored at: https://loopstransformer.netlify.app/.