Dynamic facial expression recognition (DFER) is essential to the development of intelligent and empathetic machines. Prior efforts in this field mainly fall into supervised learning paradigm, which is restricted by the limited labeled data in existing datasets. Inspired by recent unprecedented success of masked autoencoders (e.g., VideoMAE), this paper proposes MAE-DFER, a novel self-supervised method which leverages large-scale self-supervised pre-training on abundant unlabeled data to advance the development of DFER. Since the vanilla Vision Transformer (ViT) employed in VideoMAE requires substantial computation during fine-tuning, MAE-DFER develops an efficient local-global interaction Transformer (LGI-Former) as the encoder. LGI-Former first constrains self-attention in local spatiotemporal regions and then utilizes a small set of learnable representative tokens to achieve efficient local-global information exchange, thus avoiding the expensive computation of global space-time self-attention in ViT. Moreover, in addition to the standalone appearance content reconstruction in VideoMAE, MAE-DFER also introduces explicit facial motion modeling to encourage LGI-Former to excavate both static appearance and dynamic motion information. Extensive experiments on six datasets show that MAE-DFER consistently outperforms state-of-the-art supervised methods by significant margins, verifying that it can learn powerful dynamic facial representations via large-scale self-supervised pre-training. Besides, it has comparable or even better performance than VideoMAE, while largely reducing the computational cost (about 38\% FLOPs). We believe MAE-DFER has paved a new way for the advancement of DFER and can inspire more relavant research in this field and even other related tasks. Codes and models are publicly available at https://github.com/sunlicai/MAE-DFER.
Generating realistic human motion from given action descriptions has experienced significant advancements because of the emerging requirement of digital humans. While recent works have achieved impressive results in generating motion directly from textual action descriptions, they often support only a single modality of the control signal, which limits their application in the real digital human industry. This paper presents a Motion General-Purpose generaTor (MotionGPT) that can use multimodal control signals, e.g., text and single-frame poses, for generating consecutive human motions by treating multimodal signals as special input tokens in large language models (LLMs). Specifically, we first quantize multimodal control signals into discrete codes and then formulate them in a unified prompt instruction to ask the LLMs to generate the motion answer. Our MotionGPT demonstrates a unified human motion generation model with multimodal control signals by tuning a mere 0.4% of LLM parameters. To the best of our knowledge, MotionGPT is the first method to generate human motion by multimodal control signals, which we hope can shed light on this new direction. Codes shall be released upon acceptance.
Transformer is popular in recent 3D human pose estimation, which utilizes long-term modeling to lift 2D keypoints into the 3D space. However, current transformer-based methods do not fully exploit the prior knowledge of the human skeleton provided by the kinematic structure. In this paper, we propose a novel transformer-based model EvoPose to introduce the human body prior knowledge for 3D human pose estimation effectively. Specifically, a Structural Priors Representation (SPR) module represents human priors as structural features carrying rich body patterns, e.g. joint relationships. The structural features are interacted with 2D pose sequences and help the model to achieve more informative spatiotemporal features. Moreover, a Recursive Refinement (RR) module is applied to refine the 3D pose outputs by utilizing estimated results and further injects human priors simultaneously. Extensive experiments demonstrate the effectiveness of EvoPose which achieves a new state of the art on two most popular benchmarks, Human3.6M and MPI-INF-3DHP.
This technical report presents our solution to Ball Action Spotting in videos. Our method reached second place in the CVPR'23 SoccerNet Challenge. Details of this challenge can be found at https://www.soccer-net.org/tasks/ball-action-spotting. Our approach is developed based on a baseline model termed E2E-Spot, which was provided by the organizer of this competition. We first generated several variants of the E2E-Spot model, resulting in a candidate model set. We then proposed a strategy for selecting appropriate model members from this set and assigning an appropriate weight to each model. The aim of this strategy is to boost the performance of the resulting model ensemble. Therefore, we call our approach Boosted Model Ensembling (BME). Our code is available at https://github.com/ZJLAB-AMMI/E2E-Spot-MBS.
With rich visual data, such as images, becoming readily associated with items, visually-aware recommendation systems (VARS) have been widely used in different applications. Recent studies have shown that VARS are vulnerable to item-image adversarial attacks, which add human-imperceptible perturbations to the clean images associated with those items. Attacks on VARS pose new security challenges to a wide range of applications such as e-Commerce and social networks where VARS are widely used. How to secure VARS from such adversarial attacks becomes a critical problem. Currently, there is still a lack of systematic study on how to design secure defense strategies against visual attacks on VARS. In this paper, we attempt to fill this gap by proposing an adversarial image reconstruction and detection framework to secure VARS. Our proposed method can simultaneously (1) secure VARS from adversarial attacks characterized by local perturbations by image reconstruction based on global vision transformers; and (2) accurately detect adversarial examples using a novel contrastive learning approach. Meanwhile, our framework is designed to be used as both a filter and a detector so that they can be jointly trained to improve the flexibility of our defense strategy to a variety of attacks and VARS models. We have conducted extensive experimental studies with two popular attack methods (FGSM and PGD). Our experimental results on two real-world datasets show that our defense strategy against visual attacks is effective and outperforms existing methods on different attacks. Moreover, our method can detect adversarial examples with high accuracy.
Large language models (LLMs) encode a vast amount of world knowledge acquired from massive text datasets. Recent studies have demonstrated that LLMs can assist an agent in solving complex sequential decision making tasks in embodied environments by providing high-level instructions. However, interacting with LLMs can be time-consuming, as in many practical scenarios, they require a significant amount of storage space that can only be deployed on remote cloud server nodes. Additionally, using commercial LLMs can be costly since they may charge based on usage frequency. In this paper, we explore how to enable intelligent cost-effective interactions between the agent and an LLM. We propose a reinforcement learning based mediator model that determines when it is necessary to consult LLMs for high-level instructions to accomplish a target task. Experiments on 4 MiniGrid environments that entail planning sub-goals demonstrate that our method can learn to solve target tasks with only a few necessary interactions with an LLM, significantly reducing interaction costs in testing environments, compared with baseline methods. Experimental results also suggest that by learning a mediator model to interact with the LLM, the agent's performance becomes more robust against partial observability of the environment. Our code is available at https://github.com/ZJLAB-AMMI/LLM4RL.
Popularity bias is a persistent issue associated with recommendation systems, posing challenges to both fairness and efficiency. Existing literature widely acknowledges that reducing popularity bias often requires sacrificing recommendation accuracy. In this paper, we challenge this commonly held belief. Our analysis under general bias-variance decomposition framework shows that reducing bias can actually lead to improved model performance under certain conditions. To achieve this win-win situation, we propose to intervene in model training through negative sampling thereby modifying model predictions. Specifically, we provide an optimal negative sampling rule that maximizes partial AUC to preserve the accuracy of any given model, while correcting sample information and prior information to reduce popularity bias in a flexible and principled way. Our experimental results on real-world datasets demonstrate the superiority of our approach in improving recommendation performance and reducing popularity bias.
This technical report presents our Restormer-Plus approach, which was submitted to the GT-RAIN Challenge (CVPR 2023 UG$^2$+ Track 3). Details regarding the challenge are available at http://cvpr2023.ug2challenge.org/track3.html. Restormer-Plus outperformed all other submitted solutions in terms of peak signal-to-noise ratio (PSNR), and ranked 4th in terms of structural similarity (SSIM). It was officially evaluated by the competition organizers as a runner-up solution. It consists of four main modules: the single-image de-raining module (Restormer-X), the median filtering module, the weighted averaging module, and the post-processing module. Restormer-X is applied to each rainy image and built on top of Restormer. The median filtering module is used as a median operator for rainy images associated with each scene. The weighted averaging module combines the median filtering results with those of Restormer-X to alleviate overfitting caused by using only Restormer-X. Finally, the post-processing module is utilized to improve the brightness restoration. These modules make Restormer-Plus one of the state-of-the-art solutions for the GT-RAIN Challenge. Our code can be found at https://github.com/ZJLAB-AMMI/Restormer-Plus.
As Deepfake contents continue to proliferate on the internet, advancing face manipulation forensics has become a pressing issue. To combat this emerging threat, previous methods mainly focus on studying how to distinguish authentic and manipulated face images. Despite impressive, image-level classification lacks explainability and is limited to some specific application scenarios. Existing forgery localization methods suffer from imprecise and inconsistent pixel-level annotations. To alleviate these problems, this paper first re-constructs the FaceForensics++ dataset by introducing pixel-level annotations, then builds an extensive benchmark for localizing tampered regions. Next, a novel Multi-Spectral Class Center Network (MSCCNet) is proposed for face manipulation detection and localization. Specifically, inspired by the power of frequency-related forgery traces, we design Multi-Spectral Class Center (MSCC) module to learn more generalizable and semantic-agnostic features. Based on the features of different frequency bands, the MSCC module collects multispectral class centers and computes pixel-to-class relations. Applying multi-spectral class-level representations suppresses the semantic information of the visual concepts, which is insensitive to manipulations. Furthermore, we propose a Multi-level Features Aggregation (MFA) module to employ more low-level forgery artifacts and structure textures. Experimental results quantitatively and qualitatively indicate the effectiveness and superiority of the proposed MSCCNet on comprehensive localization benchmarks. We expect this work to inspire more studies on pixel-level face manipulation localization. The annotations and code will be available.
This technical report presents our Restormer-Plus approach, which was submitted to the GT-RAIN Challenge (CVPR 2023 UG$^2$+ Track 3). Details regarding the challenge are available at http://cvpr2023.ug2challenge.org/track3.html. Our Restormer-Plus outperformed all other submitted solutions in terms of peak signal-to-noise ratio (PSNR). It consists mainly of four modules: the single image de-raining module, the median filtering module, the weighted averaging module, and the post-processing module. We named the single-image de-raining module Restormer-X, which is built on Restormer and performed on each rainy image. The median filtering module is employed as a median operator for the 300 rainy images associated with each scene. The weighted averaging module combines the median filtering results with that of Restormer-X to alleviate overfitting if we only use Restormer-X. Finally, the post-processing module is used to improve the brightness restoration. Together, these modules render Restormer-Plus to be one state-of-the-art solution to the GT-RAIN Challenge. Our code is available at https://github.com/ZJLAB-AMMI/Restormer-Plus.