ACM MM 2024 Papers

Few-Shot Class-Incremental Learning has shown remarkable efficacy in efficient learning new concepts with limited annotations. Nevertheless, the heuristic few-shot annotations may not always cover the most informative samples, which largely restricts the capability of incremental learner. We aim to start from a pool of large-scale unlabeled data and then annotate the most informative samples for incremental learning. Based on this premise, Based on this purpose, this paper introduces the Active Class-Incremental Learning (ACIL). The objective of ACIL is to select the most informative samples from the unlabeled pool to effectively train an incremental learner, aiming to maximize the performance of the resulting model. Note that vanilla active learning algorithms suffer from class-imbalanced distribution among annotated samples, which restricts the ability of incremental learning. To achieve both class balance and informativeness in chosen samples, we propose Class-Balanced Selection (CBS) strategy. Specifically, we first cluster the features of all unlabeled images into multiple groups. Then for each cluster, we employ greedy selection strategy to ensure that the Gaussian distribution of the sampled features closely matches the Gaussian distribution of all unlabeled features within the cluster.Our CBS can be plugged and played into those CIL methods which are based on pretrained models with prompts tunning technique.Extensive experiments under ACIL protocol across five diverse datasets demonstrate that CBS outperforms both random selection and other SOTA active learning approaches.

Understanding human assessment of semantically salient parts of multimedia content is crucial for developing human-centric applications, such as annotation tools, search and recommender systems, and systems able to generate new media matching human interests. However, the challenge of acquiring suitable supervision signals to detect semantic saliency without extensive manual annotation remains significant. Here, we explore a novel method that utilizes signals measured directly from human cognition via electroencephalogram (EEG) in response to natural visual perception. These signals are used for supervising representation learning to capture semantic saliency. Through a contrastive learning framework, our method aligns EEG data with visual stimuli, capturing human cognitive responses without the need for any manual annotation. Our approach demonstrates that the learned representations closely align with human-centric notions of visual saliency and achieve competitive performance in several downstream tasks. We also introduce an open EEG/image dataset to facilitate research in utilizing cognitive signals for multimodal data analysis and developing models for cross-modal representation learning.

Visual abductive reasoning aims to make likely explanations for visual observations. We propose a simple yet effective Region Conditioned Adaptation, a hybrid parameter-efficient fine-tuning method that equips the frozen CLIP with the ability to infer explanations from local visual cues. We encode "local hints" and "global contexts" into visual prompts of the CLIP model separately at fine and coarse-grained levels. Adapters are used for fine-tuning CLIP models for downstream tasks and we design a new attention adapter, that directly steers the focus of the attention map with trainable query and key projections of a frozen CLIP model. Finally, we train our new model with a modified contrastive loss to regress the visual feature simultaneously toward features of literal description and plausible explanations. The loss enables CLIP to maintain both perception and reasoning abilities. Experiments on the Sherlock visual abductive reasoning benchmark show that the RCA significantly outstands previous SOTAs, ranking the 1st on the leaderboards (e.g., Human Acc: RCA 31.74 $\textit{vs}$ CPT-CLIP 29.58, higher =better). We also validate the RCA is generalizable to local perception benchmarks like RefCOCO. We open-source our project at https://github.com/LUNAProject22/RPA.

Diffusion models revolutionize image generation by leveraging natural language to guide the creation of multimedia content. Despite significant advancements in such generative models, challenges persist in depicting detailed human-object interactions, especially regarding pose and object placement accuracy. We introduce a training-free method named Reasoning and Correcting Diffusion (ReCorD) to address these challenges. Our model couples Latent Diffusion Models with Visual Language Models to refine the generation process, ensuring precise depictions of HOIs. We propose an interaction-aware reasoning module to improve the interpretation of the interaction, along with an interaction correcting module to refine the output image for more precise HOI generation delicately. Through a meticulous process of pose selection and object positioning, ReCorD achieves superior fidelity in generated images while efficiently reducing computational requirements. We conduct comprehensive experiments on three benchmarks to demonstrate the significant progress in solving text-to-image generation tasks, showcasing ReCorD's ability to render complex interactions accurately by outperforming existing methods in HOI classification score, as well as FID and Verb CLIP-Score. Project website is available at https://alberthkyhky.github.io/ReCorD/ .

Trailer generation is a challenging video clipping task that aims to select highlighting shots from long videos like movies and re-organize them in an attractive way. In this study, we propose an inverse partial optimal transport (IPOT) framework to achieve music-guided movie trailer generation. In particular, we formulate the trailer generation task as selecting and sorting key movie shots based on audio shots, which involves matching the latent representations across visual and acoustic modalities. We learn a multi-modal latent representation model in the proposed IPOT framework to achieve this aim. In this framework, a two-tower encoder derives the latent representations of movie and music shots, respectively, and an attention-assisted Sinkhorn matching network parameterizes the grounding distance between the shots' latent representations and the distribution of the movie shots. Taking the correspondence between the movie shots and its trailer music shots as the observed optimal transport plan defined on the grounding distances, we learn the model by solving an inverse partial optimal transport problem, leading to a bi-level optimization strategy. We collect real-world movies and their trailers to construct a dataset with abundant label information called CMTD and, accordingly, train and evaluate various automatic trailer generators. Compared with state-of-the-art methods, our IPOT method consistently shows superiority in subjective visual effects and objective quantitative measurements. The code is available at https://github.com/Dixin-Lab/Automatic-Movie-Trailer-Generator.

DreamBooth has demonstrated significant potential in subject-driven text-to-image generation, especially in scenarios requiring precise preservation of a subject's appearance. However, it still suffers from inefficiency and requires extensive iterative training to customize concepts using a small set of reference images. To address these issues, we introduce DreamBooth++, a region-level training strategy designed to significantly improve the efficiency and effectiveness of learning specific subjects. In particular, our approach employs a region-level data re-formulation technique that packs a set of reference images into a single sample, significantly reducing computational costs. Moreover, we adapt convolution and self-attention layers to ensure their processings are restricted within individual regions. Thus their operational scope (i.e., receptive field) can be preserved within a single subject, avoiding generating multiple sub-images within a single image. Last but not least, we design a text-guided prior regularization between our model and the pretrained one to preserve the original semantic generation ability. Comprehensive experiments demonstrate that our training strategy not only accelerates the subject-learning process but also significantly boosts fidelity to both subject and prompts in subject-driven generation.

Simultaneous Localization and Mapping (SLAM) plays a pivotal role in autonomous driving and robotics. Existing methods often rely on hand-craft feature extraction and cross-modal fusion techniques, resulting in limited feature representation capability and reduced robustness. To address this challenge, we introduce DeepPointMap2, a novel learning-based LiDAR-Visual SLAM architecture that leverages neural descriptors to tackle multiple SLAM sub-tasks in a unified manner. Our approach employs neural networks to extract multi-modal tokens, which are then adaptively fused by the Visual-Point Fusion Module to generate sparse 3D neural descriptors, ensuring precise and robust performance. As a pioneering work, our method achieves state-of-the-art localization performance among various Visual-, LiDAR-, and Visual-LiDAR-based methods in widely-used benchmarks, as shown in the experiment results. Furthermore, the approach proves to be robust in scenarios involving camera failure and LiDAR obstruction.