Open-set face forgery detection poses significant security threats and presents substantial challenges for existing detection models. These detectors primarily have two limitations: they cannot generalize across unknown forgery domains or inefficiently adapt to new data. To address these issues, we introduce an approach that is both general and parameter-efficient for face forgery detection. Our method builds on the assumption that different forgery source domains exhibit distinct style statistics. Specifically, we design a forgery-style-mixture formulation that augments the diversity of forgery source domains, enhancing the model’s generalizability across unseen domains. In addition, previous methods typically require fully fine-tuning pretrained networks, consuming substantial time and computational resources. Drawing on recent advancements in vision transformers (ViT) for face forgery detection, we develop a parameter-efficient ViT-based detection model that includes lightweight forgery feature extraction modules and enables the model to extract global and local forgery clues simultaneously. We only optimize the inserted lightweight modules during training, maintaining the original ViT structure with its pre-trained weights. This training strategy effectively preserves the informative pre-trained knowledge while flexibly adapting the model to the task of Deepfake detection. Extensive experimental results demonstrate that the designed model achieves state-of-the-art generalizability with significantly reduced trainable parameters, representing an important step toward open-set Deepfake detection in the wild.

Deepfakes have recently raised significant trust issues and security concerns among the public. Compared to CNN-based face forgery detectors, ViT-based methods take advantage of the expressivity of transformers, achieving superior detection performance. However, these approaches still exhibit the following limitations: (1) Fully fine-tuning ViT-based models from ImageNet weights demands substantial computational and storage resources; (2) ViT-based methods struggle to capture local forgery clues, leading to model bias; (3) These methods limit their scope on only one or few face forgery features, resulting in limited generalizability. To tackle these challenges, this work introduces Mixture-of-Experts modules for Face Forgery Detection (MoE-FFD), a generalized yet parameter-efficient ViT-based approach. MoE-FFD only updates lightweight Low-Rank Adaptation (LoRA) and Adapter layers while keeping the ViT backbone frozen, thereby achieving parameter-efficient training. Moreover, MoE-FFD leverages the expressivity of transformers and local priors of CNNs to simultaneously extract global and local forgery clues. Additionally, novel MoE modules are designed to scale the model's capacity and smartly select optimal forgery experts, further enhancing forgery detection performance. Our proposed learning scheme can be seamlessly adapted to various transformer backbones in a plug-and-play manner. Extensive experimental results demonstrate that the proposed method achieves state-of-the-art face forgery detection performance with significantly reduced parameter overhead in cross-dataset, cross-manipulation, and robustness evaluations. Our ablation studies further validate the effectiveness of the designed components and the proposed learning scheme. The code is available at: https://github.com/LoveSiameseCat/MoE-FFD.

In this paper, we tackle the open-set temporal action segmentation task, which aims to identify unknown frames while ensuring accurate segmentation of known actions in the temporal domain. Existing open-set methods struggle with identifying unknown frames due to their indistinguishability against ambiguous known frames during action transitions, resulting in significant performance degradation. To address this, we propose the action distribution flow, which models transitions between action sequences to capture the inherent feature discrepancies between unknown and known frames. Specifically, our method first models the distributions of known actions using the training data, and then interpolates these distributions along the optimal transport path for consecutive actions in the testing videos. By evaluating the likelihood of testing frames against the modeled action distribution flow, our approach effectively identifies unknown frames without requiring additional training or prior knowledge of the unknown data. Extensive experiments on open-set versions of the GTEA, 50Salads, and Breakfast datasets demonstrate the superiority of the proposed method across all evaluation metrics.

Natural language video localization plays a pivotal role in video understanding, and leveraging weakly-labeled data is considered a promising approach to circumvent the laborintensive process of manual annotations. However, this approach encounters two significant challenges: 1) limited input distribution, namely that the limited writing styles of the language query, annotated by human annotators, hinder the model’s generalization to real-world scenarios with diverse vocabularies and sentence structures; 2) the incomplete ground truth, whose supervision guidance is insufficient. To overcome these challenges, we propose an omnipotent distillation algorithm with large language models (LLM). The distribution of the input sample is enriched to obtain diverse multi-view versions while a consistency then comes to regularize the consistency of their results for distillation. Specifically, we first train our teacher model with the proposed intra-model agreement, where multiple sub-models are supervised by each other. Then, we leverage the LLM to paraphrase the language query and distill the teacher model to a lightweight student model by enforcing the consistency between the localization results of the paraphrased sentence and the original one. In addition, to assess the generalization of the model across different dimensions of language variation, we create extensive datasets by building upon existing datasets. Our experiments demonstrate substantial performance improvements adaptively to diverse kinds of language queries.

This paper for the first time leverages multi-modal videos for weakly-supervised temporal video grounding. As labeling the video moment is labor-intensive and subjective, the weakly-supervised approaches have gained increasing attention in recent years. However, these approaches could inherently compromise performance due to inadequate supervision. Therefore, to tackle this challenge, we for the first time pay attention to exploiting complementary information extracted from multi-modal videos (e.g., RGB frames, optical flows), where richer supervision is naturally introduced in the weaklysupervised context. Our motivation is that by integrating different modalities of the videos, the model is learned from synergic supervision and thereby can attain superior generalization capability. However, addressing multiple modalities† would also inevitably introduce additional computational overhead, and might become inapplicable if a particular modality is inaccessible. To solve this issue, we adopt a novel route: building a multi-modal distillation algorithm to capitalize on the multi-modal knowledge as supervision for model training, while still being able to work with only the single modal input during inference. As such, we can utilize the benefits brought by the supplementary nature of multiple modalities, without compromising the applicability in practical scenarios. Specifically, we first propose a cross-modal mutual learning framework and train a sophisticated teacher model to learn collaboratively from the multi-modal videos. Then we identify two sorts of knowledge from the teacher model, i.e., temporal boundaries and semantic activation map. And we devise a local-global distillation algorithm to transfer this knowledge to a student model of single-modal input at both local and global levels. Extensive experiments on large-scale datasets demonstrate that our method achieves state-of-the-art performance with/without multi-modal inputs.

This paper for the first time explores audio-visual event localization in an unsupervised manner. Previous methods tackle this problem in a supervised setting and require segment-level or video-level event category ground-truth to train the model. However, building large-scale multi-modality datasets with category annotations is human-intensive and thus not scalable to real-world applications. To this end, we propose cross-modal label contrastive learning to exploit multi-modal information among unlabeled audio and visual streams as self-supervision signals. At the feature representation level, multi-modal representations are collaboratively learned from audio and visual components by using self-supervised representation learning. At the label level, we propose a novel self-supervised pretext task i.e. label contrasting to self-annotate videos with pseudo-labels for localization model training. Note that irrelevant background would hinder the acquisition of high-quality pseudo-labels and thus lead to an inferior localization model. To address this issue, we then propose an expectation-maximization algorithm that optimizes the pseudo-label acquisition and localization model in a coarse-to-fine manner. Extensive experiments demonstrate that our unsupervised approach performs reasonably well compared to the state-of-the-art supervised methods.

The task of temporal grounding aims to locate video moment in an untrimmed video, with a given sentence query. This paper for the first time investigates some superficial biases that are specific to the temporal grounding task, and proposes a novel targeted solution. Most alarmingly, we observe that existing temporal ground models heavily rely on some biases (e.g., high preference on frequent concepts or certain temporal intervals) in the visual modal. This leads to inferior performance when generalizing the model in cross-scenario test setting. To this end, we propose a novel method called Debiased Temporal Language Localizer (Debias-TLL) to prevent the model from naively memorizing the biases and enforce it to ground the query sentence based on true inter-modal relationship. Debias-TLL simultaneously trains two models. By our design, a large discrepancy of these two models' predictions when judging a sample reveals higher probability of being a biased sample. Harnessing the informative discrepancy, we devise a data re-weighing scheme for mitigating the data biases. We evaluate the proposed model in cross-scenario temporal grounding, where the train / test data are heterogeneously sourced. Experiments show large-margin superiority of the proposed method in comparison with state-of-the-art competitors.

This paper explores a novel setting of temporal sentence grounding for the first time, dubbed as dense events grounding. Given an untrimmed video and a paragraph description, dense events grounding aims to jointly localize temporal moments of multiple events described in the paragraph. Our main motivating fact is that multiple events to be grounded in a video are often semantically related and temporally coordinated according to their order appearing in the paragraph. This fact sheds light on devising more accurate visual grounding model. In this work, we propose Dense Events Propagation Network (DepNet) for this novel task. DepNet first adaptively aggregates temporal and semantic information of dense events into a compact set through a second-order attention pooling, then selectively propagates the aggregated information to each single event with soft attention. Based on such aggregation-and-propagation mechanism, DepNet can effectively exploit both the temporal order and semantic relations of dense events. We conduct comprehensive experiments on large-scale datasets ActivityNet Captions and TACoS. For fair comparisons, our evaluations include both state-of-art single-event grounding methods and their natural extensions to the dense-events grounding setting implemented by us. All experiments clearly shows the performance superiority of the proposed DepNet by significant margins.

3-D human pose estimation is a crucial step for understanding human actions. However, reliably capturing precise 3-D position of human joints is non-trivial and tedious. Current models often suffer from the scarcity of high-quality 3-D annotated training data. In this work, we explore a novel way of obtaining gigantic 3-D human pose data without manual annotations. In catedioptric videos (\emph{e.g.}, people dance before a mirror), the camera records both the original and mirrored human poses, which provides cues for estimating 3-D positions of human joints. Following this idea, we crawl a large-scale Dance-before-Mirror (DBM) video dataset, which is about 24 times larger than existing Human3.6M benchmark. Our technical insight is that, by jointly harnessing the epipolar geometry and human skeleton priors, 3-D joint estimation can boil down to an optimization problem over two sets of 2-D estimations. To our best knowledge, this represents the first work that collects high-quality 3-D human data via catadioptric systems. We have conducted comprehensive experiments on cross-scenario pose estimation and visualization analysis. The results strongly demonstrate the usefulness of our proposed DBM human poses.