ACM MM 2024 Papers

Multi-modal Unsupervised Domain Adaptation (MM-UDA) for large-scale 3D semantic segmentation involves adapting 2D and 3D models to a target domain without labels, which significantly reduces the labor-intensive annotations. Existing MM-UDA methods have often attempted to mitigate the domain discrepancy by aligning features between the source and target data. However, this implementation falls short when applied to image perception due to the susceptibility of images to environmental changes compared to point clouds. To mitigate this limitation, in this work, we explore the potentials of an off-the-shelf Contrastive Language-Image Pre-training (CLIP) model with rich whilst heterogeneous knowledge. To make CLIP task-specific, we propose a top-performing method, dubbed CLIP2UDA, which makes frozen CLIP reward unsupervised domain adaptation in 3D semantic segmentation. Specifically, CLIP2UDA alternates between two steps during adaptation: (a) Learning task-specific prompt. 2D features response from the visual encoder are employed to initiate the learning of adaptive text prompt of each domain, and (b) Learning multi-modal domain-invariant representations. These representations interact hierarchically in the shared decoder to obtain unified 2D visual predictions. This enhancement allows for effective alignment between the modality-specific 3D and unified feature space via cross-modal mutual learning. Extensive experimental results demonstrate that our method outperforms state-of-the-art competitors in several widely-recognized adaptation scenarios. Code is available at: https://github.com/Barcaaaa/CLIP2UDA.

In e-commerce platforms, visual content plays a pivotal role in capturing and retaining audience attention. A high-quality and aesthetically designed product background image can quickly grab consumers' attention, and increase their confidence in taking actions, such as making a purchase. Recently, diffusion models have achieved profound advancements, rendering product background generation a promising avenue for exploration. However, text-guided diffusion models require meticulously crafted prompts. The diverse range of products makes it challenging to compose prompts that result in visually appealing and semantically appropriate background scenes. Current work has made great efforts on creating prompts through expert-crafted rules or specialized fine-tuning of large language models, but it still relies on detailed human inputs and often falls short in generating desirable results by e-commerce standards.

Although current prompt learning methods have successfully been designed to effectively reuse the large pre-trained models without fine-tuning their large number of parameters, they still have limitations to be addressed, i.e., without considering the adverse impact of meaningless patches in every image and without simultaneously considering in-sample generalization and out-of-sample generalization. In this paper, we propose an adaptive multi-modality prompt learning to address the above issues. To do this, we employ previous text prompt learning and propose a new image prompt learning. The image prompt learning achieves in-sample and out-of-sample generalization, by first masking meaningless patches and then padding them with the learnable parameters and the information from texts. Moreover, each of the prompts provides auxiliary information to each other, further strengthening these two kinds of generalization. Experimental results on real datasets demonstrate that our method outperforms SOTA methods, in terms of different downstream tasks.

The core of active semi-supervised crowd counting is the sample selection criteria. However, the scale factor has been neglected in active learning approaches despite the fact that the scale of heads varies drastically in the crowd images. In this paper, we propose a simple yet effective active labeling strategy to explicitly select informative unlabeled images, guided by the intra-scale uncertainty and inter-scale inconsistency metrics. The intra-scale uncertainty is quantified through the sum of the query-level entropy of images at different scales. Images are initially ranked based on this uncertainty for preselection. Inter-scale inconsistency is measured by the divergence between the query-level predictions of upscaled and downscaled images, allowing for the identification of the most informative images exhibiting the highest inconsistency. Additionally, we implement a progressive updating scheme for the semi-supervised crowd counting framework, in which the pseudo-labels for unlabeled images are refined iteratively. It further improves the counting accuracy. Through extensive experiments on widely used benchmarks, the proposed approach has demonstrated superior performance compared to previous state-of-the-art semi-supervised and active semi-supervised crowd counting methods.

Image-based 3D Virtual Try-ON (VTON) aims to sculpt the 3D human according to person and clothes images, which is data-efficient (i.e., getting rid of expensive 3D data) but challenging. Recent text-to-3D methods achieve remarkable improvement in high-fidelity 3D human generation, demonstrating its potential for 3D virtual try-on. Inspired by the impressive success of personalized diffusion models (e.g., Dreambooth and LoRA) for 2D VTON, it is straightforward to achieve 3D VTON by integrating the personalization technique into the diffusion-based text-to-3D framework. However, employing the personalized module in a pre-trained diffusion model (e.g., StableDiffusion (SD)) would degrade the model's capability for multi-view or multi-domain synthesis, which is detrimental to the geometry and texture optimization guided by Score Distillation Sampling (SDS) loss. In this work, we propose a novel customizing 3D human try-on model, named \textbf{DreamVTON}, to separately optimize the geometry and texture of the 3D human. Specifically, a personalized SD with multi-concept LoRA is proposed to provide the generative prior about the specific person and clothes, while a Densepose-guided ControlNet is exploited to guarantee consistent prior about body pose across various camera views. Besides, to avoid the inconsistent multi-view priors from the personalized SD dominating the optimization, DreamVTON introduces a template-based optimization mechanism, which employs mask templates for geometry shape learning and normal/RGB templates for geometry/texture details learning. Furthermore, for the geometry optimization phase, DreamVTON integrates a normal-style LoRA into personalized SD to enhance normal map generative prior, facilitating smooth geometry modeling.

Test-time adaptation (TTA) aims to adapt a model, initially trained on training data, to test data with potential distribution shifts. Most existing TTA methods focus on classification problems. The pronounced success of classification might lead numerous newcomers and engineers to assume that classic TTA techniques can be directly applied to the more challenging task of semantic segmentation. However, this belief is still an open question. In this paper, we investigate the applicability of existing classic TTA strategies in semantic segmentation. Our comprehensive results have led to three key observations. First, the classic normalization updating strategy only brings slight performance improvement, and in some cases, it might even adversely affect the results. Even with the application of advanced distribution estimation techniques like batch renormalization, the problem remains unresolved. Second, although the teacher-student scheme does enhance the training stability for segmentation TTA in the presence of noisy pseudo-labels and temporal correlation, it cannot directly result in performance improvement compared to the original model without TTA under complex data distribution. Third, segmentation TTA suffers a severe long-tailed class-imbalance problem, which is substantially more complex than that in TTA for classification. This long-tailed challenge negatively affects segmentation TTA performance, even when the accuracy of pseudo-labels is high. Besides those observations, we find that visual prompt tuning (VisPT) is promising in segmentation TTA and propose a novel method named TTAP. The outstanding performance of TTAP has also been verified. We hope the community can give more attention to this challenging, yet important, segmentation TTA task in the future. The source code is available at: https://github.com/ycarobot/TTAP.

Reconstructing 3D scenes from multi-view images is challenging, especially under extreme scenarios. We propose Event-ID, an event-based intrinsic decomposition framework that leverages events and images for stable decomposition under extreme scenarios. Our method is based on two observations: event cameras maintain good imaging quality under blurry or poorly exposed scenarios, and event signals from different viewpoints exhibit similarity in diffuse regions while varying in specular regions. We establish an event-based reflectance model and introduce an event-based warping method to extract specular clues. Our two-stage framework constructs a radiance field and decomposes the scene into normal, material, and lighting. Experimental results demonstrate superior performance compared to state-of-the-art methods. Our project can be found at https://zehaoc.github.io/EventID.github.io/

Most existing methods for text-based person retrieval focus on text-to-image person retrieval. Nevertheless, due to the lack of dynamic information provided by isolated frames, the performance is hampered when the person is obscured or variable motion details are missed in isolated frames. To overcome this, we propose a novel Text-to-Video Person Retrieval (TVPR) task. Since there is no dataset or benchmark that describes person videos with natural language, we construct a large-scale cross-modal person video dataset containing detailed natural language annotations, termed as Text-to-Video Person Reidentification (TVPReid) dataset. In this paper, we introduce a Multielement Feature Guided Fragments Learning (MFGF) strategy, which leverages the cross-modal text-video representations to provide strong text-visual and text-motion matching information to tackle uncertain occlusion conflicting and variable motion details. Specifically, we establish two potential cross-modal spaces for text and video feature collaborative learning to progressively reduce the semantic difference between text and video. To evaluate the effectiveness of the proposed MFGF, extensive experiments have been conducted on TVPReid dataset. To the best of our knowledge, MFGF is the first successful attempt to use video for text-based person retrieval task and has achieved state-of-the-art performance on TVPReid dataset. The TVPReid dataset will be publicly available to benefit future research.

Text to Motion Retrieval (TMR) is an emerging task to retrieve relevant motion sequences with the nature language description. The dominant approach learns a joint embedding space to measure global-level similarities. However, simple global embeddings are insufficient to represent complicated motion and textual details, such as the movement of specific body parts and the coordination among these body parts. In addition, most of the motion variations occur subtly and locally, resulting in semantic vagueness among these motions, which further presents considerable challenges in precisely aligning motion sequences with texts. To address these challenges, we propose a novel Modal-Enhanced Semantic Modeling (MESM) method, focusing on fine-grained alignment through enhanced modal semantics. Specifically, we develop a prompt-enhanced textual module (PTM) to generate detailed descriptions of specific body part movements, which comprehensively captures the fine-grained textual semantics for precise matching. We employ a skeleton-enhanced motion module (SMM) to effectively enhance the model's capability to represent intricate motions. This module leverages a graph convolutional network to meticulously model the intricate spatial dependencies among relevant body parts. To improve the sensitivity to the subtle motions, we further propose a text-driven semantics interaction module (TSIM). The TSIM assigns motion features into a set of aggregated descriptors and employs cross-attention to aggregate discriminative motion embeddings guided by text, enabling precise semantic alignment between subtle motions and corresponding texts. Extensive experiments conducted on two widely used benchmark datasets, HumanML3D and KIT-ML, demonstrate the effectiveness of our proposed method. Our approach outperforms existing state-of-the-art retrieval methods, achieving significant Rsum improvements of 24.28% on HumanML3D and 25.80% on KIT-ML.

With the rise of Machine Learning as a Service (MLaaS) platforms, safeguarding the intellectual property of deep learning models is becoming paramount. Among various protective measures, trigger set watermarking has emerged as a flexible and effective strategy for preventing unauthorized model distribution. However, this paper identifies an inherent flaw in the current paradigm of trigger set watermarking: evasion adversaries can readily exploit the shortcuts created by models memorizing watermark samples that deviate from the main task distribution, significantly impairing their generalization in adversarial settings. To counteract this, we leverage diffusion models to synthesize unrestricted adversarial examples as trigger sets. By learning the model to accurately recognize them, unique watermark behaviors are promoted through knowledge injection rather than error memorization, thus avoiding exploitable shortcuts. Furthermore, we uncover that the resistance of current trigger set watermarking against removal attacks primarily relies on significantly damaging the decision boundaries during embedding, intertwining unremovability with adverse impacts. By optimizing the knowledge transfer properties of protected models, our approach conveys watermark behaviors to extraction surrogates without aggressive decision boundary perturbation. Experimental results on CIFAR-10/100 and Imagenette datasets demonstrate the effectiveness of our method, showing not only improved robustness against evasion adversaries but also superior resistance to watermark removal attacks compared to state-of-the-art solutions.

In recent years, Few-Shot Object Detection (FSOD) has gained widespread attention and made significant progress due to its ability to build models with a good generalization power using extremely limited annotated data. The fine-tuning based paradigm is currently dominating this field, where detectors are initially pre-trained on base classes with sufficient samples and then fine-tuned on novel ones with few samples, but the scarcity of labeled samples of novel classes greatly interferes precisely fitting their data distribution, thus hampering the performance. To address this issue, we propose a new framework for FSOD, namely Prototype-based Soft-labels and Test-Time Learning (PS-TTL). Specifically, we design a Test-Time Learning (TTL) module that employs a mean-teacher network for self-training to discover novel instances from test data, allowing detectors to learn better representations and classifiers for novel classes. Furthermore, we notice that even though relatively low-confidence pseudo-labels exhibit classification confusion, they still tend to recall foreground. We thus develop a Prototype-based Soft-labels (PS) strategy through assessing similarities between low-confidence pseudo-labels and category prototypes as soft-labels to unleash their potential, which substantially mitigates the constraints posed by few-shot samples. Extensive experiments on both the VOC and COCO benchmarks show that PS-TTL achieves the state-of-the-art, highlighting its effectiveness. The code and model are available at https://github.com/gaoyingjay/PS-TTL.

More and more end-to-end text spotting methods based on Transformer architecture have demonstrated superior performance. These methods utilize a bipartite graph matching algorithm to perform one-to-one optimal matching between predicted objects and actual objects. However, the instability of bipartite graph matching can lead to inconsistent optimization targets, thereby affecting the training performance of the model. Existing literature applies denoising training to solve the problem of bipartite graph matching instability in object detection tasks. Unfortunately, this denoising training method cannot be directly applied to text spotting tasks, as these tasks need to perform irregular shape detection tasks and more complex text recognition tasks than classification. To address this issue, we propose a novel denoising training method (DNTextSpotter) for arbitrary-shaped text spotting. Specifically, we decompose the queries of the denoising part into noised positional and noised content queries. We use the four Bezier control points of the Bezier center curve to generate the noised positional queries. For the noised content queries, considering that the output of the text in a fixed positional order is not conducive to aligning position with content, we employ a masked character sliding method to initialize noised content queries, thereby assisting in the alignment of text content and position. Additionally, to improve the model's perception of the background, we further utilize an additional loss function for background characters classification in the denoising training part. DNTextSpotter outperforms state-of-the-art methods on four benchmarks-Total-Text, SCUT-CTW1500, ICDAR15, and Inverse-Text-most notably achieving an 11.3% improvement over the best approach on Inverse-Text.

Joint Multimodal Entity-Relation Extraction (JMERE) is a challenging task that aims to extract entities and their relations from textimage pairs in social media posts. Existing methods for JMERE require large amounts of labeled data. However, gathering and annotating fine-grained multimodal data for JMERE poses significant challenges. Initially, we construct diverse and comprehensive multimodal few-shot datasets fitted to the original data distribution. To address the insufficient information in the few-shot setting, we introduce the Knowledge-Enhanced Cross-modal Prompt Model (KECPM) for JMERE. This method can effectively address the problem of insufficient information in the few-shot setting by guiding a large language model to generate supplementary background knowledge. Our proposed method comprises two stages: (1) a knowledge ingestion stage that dynamically formulates prompts based on semantic similarity guide ChatGPT generating relevant knowledge and employs self-reflection to refine the knowledge; (2) a knowledge-enhanced language model stage that merges the auxiliary knowledge with the original input and utilizes a transformerbased model to align with JMERE's required output format. We extensively evaluate our approach on a few-shot dataset derived from the JMERE dataset, demonstrating its superiority over strong baselines in terms of both micro and macro F1 scores. Additionally, we present qualitative analyses and case studies to elucidate the effectiveness of our model. Code and Data are released at https://github.com/YuanLi95/KECPM.

In recent years, multi-view outlier detection (MVOD) methods have advanced significantly, aiming to identify outliers within multi-view datasets. A key point is to better detect class outliers and class-attribute outliers, which only exist in multi-view data. However, existing methods either is not able to reduce the impact of outliers when learning view-consistent information, or struggle in cases with varying neighborhood structures. Moreover, most of them do not apply to partial multi-view data in real-world scenarios. To overcome these drawbacks, we propose a novel method named Regularized Contrastive Partial Multi-view Outlier Detection (RCPMOD). In this framework, we utilize contrastive learning to learn view-consistent information and distinguish outliers by the degree of consistency. Specifically, we propose (1) An outlier-aware contrastive loss with a potential outlier memory bank to eliminate their bias motivated by a theoretical analysis. (2) A neighbor alignment contrastive loss to capture the view-shared local structural correlation. (3) A spreading regularization loss to prevent the model from overfitting over outliers. With the Cross-view Relation Transfer technique, we could easily impute the missing view samples based on the features of neighbors. Experimental results on four benchmark datasets demonstrate that our proposed approach could outperform state-of-the-art competitors under different settings.

Neural Radiance Fields (NeRFs) demonstrate high efficiency in generating photo-realistic novel view. Recent studies introduce the trials on the 3D inpainting by NeRF. However, the performance of these works have been validated for data collected in a narrow range of multi-view, while degrade for the wide range of multi-view. To address this problem, we propose a novel NeRF framework to remove the obstacle and reproduce occluded areas in high quality for both wide and narrow range of multi-view. In this framework, we design a region coding network to carry out object segmentation. With the depth information, the segmentation component transfers a single obstacle mask to other views in high accuracy. By referring to the segmentation results, we introduce an innovative view selection mechanism to reconstruct the occluded area using supplementary information from multi-view and 2D inpainting. We also contribute to the evaluation of 3D scene de-occlusion by introducing a dataset including views captured in wide range and in pair with and without the obstacle object for comparison. We evaluate our framework in both narrow and wide range datasets by quantitative measurement and visually qualitative comparison, which confirm the competitive and superior performance of our framework.

The massive generation of multimodal fake news involving both text and images exhibits substantial distribution discrepancies, prompting the need for generalized detectors. However, the insulated nature of training restricts the capability of classical detectors to obtain open-world facts. While Large Vision-Language Models (LVLMs) have encoded rich world knowledge, they are not inherently tailored for combating fake news and struggle to comprehend local forgery details. In this paper, we propose FKA-Owl, a novel framework that leverages forgery-specific knowledge to augment LVLMs, enabling them to reason about manipulations effectively. The augmented forgery-specific knowledge includes semantic correlation between text and images, and artifact trace in image manipulation. To inject these two kinds of knowledge into the LVLM, we design two specialized modules to establish their representations, respectively. The encoded knowledge embeddings are then incorporated into LVLMs. Extensive experiments on the public benchmark demonstrate that FKA-Owl achieves superior cross-domain performance compared to previous methods. Code is publicly available at https://liuxuannan.github.io/FKA_Owl.github.io/.

Multimodal Emotion Recognition in Conversations (MERC) aims to identify the emotions conveyed by each utterance in a conversational video. Current efforts focus on modeling speaker-sensitive context dependencies and multimodal fusion. Despite the progress, the reliability of MERC methods remains largely unexplored. Extensive empirical studies reveal that current methods suffer from unreliable predictive confidence. Specifically, in some cases, the confidence estimated by these models increases when a modality or specific contextual cues are corrupted, defining these as uncertain samples. This contradicts the foundational principle in informatics, namely, the elimination of uncertainty. Based on this, we propose a novel calibration framework CMERC to calibrate MERC models without altering the model structure. It integrates curriculum learning to guide the model in progressively learning more uncertain samples; hybrid supervised contrastive learning to refine utterance representations, by pulling uncertain samples and others apart; and confidence constraint to penalize the model on uncertain samples. Experimental results on two datasets demonstrate the effectiveness and generalization capabilities of our CMERC across various MERC models, surpassing state-of-the-art methods.

In online chatting, people increasingly prefer using stickers to supplement or replace text for replies, as sticker images can express vivid and varied emotions. The Sticker Response Selection (SRS) task aims to predict the sticker image that is most relevant to the history dialogue. Previous researches explore the semantic similarity between context and stickers, overlooking both unimodal and cross-modal emotional information. In this paper, we propose a 'Perceive before Respond' (PBR) training paradigm. PBR perceives sticker emotions through a knowledge distillation module. Variety representations of each emotion category are acquired from the large-scale sticker emotion recognition dataset and distilled into our model to enhance emotion comprehension. We further distinguish stickers with similar subject elements under the same topic. We perform contrastive learning at both inter- and intra-topic levels to obtain discriminative and diverse sticker representations. In addition, we improve the hard negative sampling method for image-text matching based on cross-modal sentiment association, conducting hard sample mining from both semantic similarity and sentiment polarity similarity for sticker-to-dialogue and dialogue-to-sticker. Extensive experiments verify the effectiveness of each proposed component. Ablation experiments on different backbone networks demonstrate the generality of our approach. Our code is released on https://github.com/wuyou-xia/Perceive-before-Respond.

Co-speech gesture generation is crucial for producing synchronized and realistic human gestures that accompany speech, enhancing the animation of lifelike avatars in virtual environments. While diffusion models have shown impressive capabilities, current approaches often overlook a wide range of modalities and their interactions, resulting in less dynamic and contextually varied gestures. To address these challenges, we present MambaGesture, a novel framework integrating a Mamba-based attention block, MambaAttn, with a multi-modality feature fusion module, SEAD. The MambaAttn block combines the sequential data processing strengths of the Mamba model with the contextual richness of attention mechanisms, enhancing the temporal coherence of generated gestures. SEAD adeptly fuses audio, text, style, and emotion modalities, employing disentanglement to deepen the fusion process and yield gestures with greater realism and diversity. Our approach, rigorously evaluated on the multi-modal BEAT dataset, demonstrates significant improvements in Fréchet Gesture Distance (FGD), diversity scores, and beat alignment, achieving state-of-the-art performance in co-speech gesture generation.