Bonan Li

Bonan Li, Yinhan Hu, Xuecheng Nie et al.

In this paper, we focus on analyzing and improving the dropout technique for self-attention layers of Vision Transformer, which is important while surprisingly ignored by prior works. In particular, we conduct researches on three core questions: First, what to drop in self-attention layers? Different from dropping attention weights in literature, we propose to move dropout operations forward ahead of attention matrix calculation and set the Key as the dropout unit, yielding a novel dropout-before-softmax scheme. We theoretically verify that this scheme helps keep both regularization and probability features of attention weights, alleviating the overfittings problem to specific patterns and enhancing the model to globally capture vital information; Second, how to schedule the drop ratio in consecutive layers? In contrast to exploit a constant drop ratio for all layers, we present a new decreasing schedule that gradually decreases the drop ratio along the stack of self-attention layers. We experimentally validate the proposed schedule can avoid overfittings in low-level features and missing in high-level semantics, thus improving the robustness and stableness of model training; Third, whether need to perform structured dropout operation as CNN? We attempt patch-based block-version of dropout operation and find that this useful trick for CNN is not essential for ViT. Given exploration on the above three questions, we present the novel DropKey method that regards Key as the drop unit and exploits decreasing schedule for drop ratio, improving ViTs in a general way. Comprehensive experiments demonstrate the effectiveness of DropKey for various ViT architectures, e.g. T2T and VOLO, as well as for various vision tasks, e.g., image classification, object detection, human-object interaction detection and human body shape recovery.

Ziwen Liu, Bonan Li, Congying Han et al.

Contrastive learning (CL) has shown great power in self-supervised learning due to its ability to capture insight correlations among large-scale data. Current CL models are biased to learn only the ability to discriminate positive and negative pairs due to the discriminative task setting. However, this bias would lead to ignoring its sufficiency for other downstream tasks, which we call the discriminative information overfitting problem. In this paper, we propose to tackle the above problems from the aspect of the Information Bottleneck (IB) principle, further pushing forward the frontier of CL. Specifically, we present a new perspective that CL is an instantiation of the IB principle, including information compression and expression. We theoretically analyze the optimal information situation and demonstrate that minimum sufficient augmentation and information-generalized representation are the optimal requirements for achieving maximum compression and generalizability to downstream tasks. Therefore, we propose the Masked Reconstruction Contrastive Learning~(MRCL) model to improve CL models. For implementation in practice, MRCL utilizes the masking operation for stronger augmentation, further eliminating redundant and noisy information. In order to alleviate the discriminative information overfitting problem effectively, we employ the reconstruction task to regularize the discriminative task. We conduct comprehensive experiments and show the superiority of the proposed model on multiple tasks, including image classification, semantic segmentation and objective detection.

Zicheng Zhang, Ruobing Zheng, Ziwen Liu et al.

Recent works in implicit representations, such as Neural Radiance Fields (NeRF), have advanced the generation of realistic and animatable head avatars from video sequences. These implicit methods are still confronted by visual artifacts and jitters, since the lack of explicit geometric constraints poses a fundamental challenge in accurately modeling complex facial deformations. In this paper, we introduce Dynamic Tetrahedra (DynTet), a novel hybrid representation that encodes explicit dynamic meshes by neural networks to ensure geometric consistency across various motions and viewpoints. DynTet is parameterized by the coordinate-based networks which learn signed distance, deformation, and material texture, anchoring the training data into a predefined tetrahedra grid. Leveraging Marching Tetrahedra, DynTet efficiently decodes textured meshes with a consistent topology, enabling fast rendering through a differentiable rasterizer and supervision via a pixel loss. To enhance training efficiency, we incorporate classical 3D Morphable Models to facilitate geometry learning and define a canonical space for simplifying texture learning. These advantages are readily achievable owing to the effective geometric representation employed in DynTet. Compared with prior works, DynTet demonstrates significant improvements in fidelity, lip synchronization, and real-time performance according to various metrics. Beyond producing stable and visually appealing synthesis videos, our method also outputs the dynamic meshes which is promising to enable many emerging applications.

Xinmin Qiu, Congying Han, Zicheng Zhang et al.

Developing blind video deflickering (BVD) algorithms to enhance video temporal consistency, is gaining importance amid the flourish of image processing and video generation. However, the intricate nature of video data complicates the training of deep learning methods, leading to high resource consumption and instability, notably under severe lighting flicker. This underscores the critical need for a compact representation beyond pixel values to advance BVD research and applications. Inspired by the classic scale-time equalization (STE), our work introduces the histogram-assisted solution, called BlazeBVD, for high-fidelity and rapid BVD. Compared with STE, which directly corrects pixel values by temporally smoothing color histograms, BlazeBVD leverages smoothed illumination histograms within STE filtering to ease the challenge of learning temporal data using neural networks. In technique, BlazeBVD begins by condensing pixel values into illumination histograms that precisely capture flickering and local exposure variations. These histograms are then smoothed to produce singular frames set, filtered illumination maps, and exposure maps. Resorting to these deflickering priors, BlazeBVD utilizes a 2D network to restore faithful and consistent texture impacted by lighting changes or localized exposure issues. BlazeBVD also incorporates a lightweight 3D network to amend slight temporal inconsistencies, avoiding the resource consumption issue. Comprehensive experiments on synthetic, real-world and generated videos, showcase the superior qualitative and quantitative results of BlazeBVD, achieving inference speeds up to 10x faster than state-of-the-arts.

Zicheng Zhang, Bonan Li, Xuecheng Nie et al.

Existing works have advanced Text-to-Image (TTI) diffusion models for video editing in a one-shot learning manner. Despite their low requirements of data and computation, these methods might produce results of unsatisfied consistency with text prompt as well as temporal sequence, limiting their applications in the real world. In this paper, we propose to address the above issues with a novel EI$^2$ model towards \textbf{E}nhancing v\textbf{I}deo \textbf{E}diting cons\textbf{I}stency of TTI-based frameworks. Specifically, we analyze and find that the inconsistent problem is caused by newly added modules into TTI models for learning temporal information. These modules lead to covariate shift in the feature space, which harms the editing capability. Thus, we design EI$^2$ to tackle the above drawbacks with two classical modules: Shift-restricted Temporal Attention Module (STAM) and Fine-coarse Frame Attention Module (FFAM). First, through theoretical analysis, we demonstrate that covariate shift is highly related to Layer Normalization, thus STAM employs a \textit{Instance Centering} layer replacing it to preserve the distribution of temporal features. In addition, {STAM} employs an attention layer with normalized mapping to transform temporal features while constraining the variance shift. As the second part, we incorporate {STAM} with a novel {FFAM}, which efficiently leverages fine-coarse spatial information of overall frames to further enhance temporal consistency. Extensive experiments demonstrate the superiority of the proposed EI$^2$ model for text-driven video editing.

Xinmin Qiu, Congying Han, Zicheng Zhang et al.

Blind face restoration (BFR) is important while challenging. Prior works prefer to exploit GAN-based frameworks to tackle this task due to the balance of quality and efficiency. However, these methods suffer from poor stability and adaptability to long-tail distribution, failing to simultaneously retain source identity and restore detail. We propose DiffBFR to introduce Diffusion Probabilistic Model (DPM) for BFR to tackle the above problem, given its superiority over GAN in aspects of avoiding training collapse and generating long-tail distribution. DiffBFR utilizes a two-step design, that first restores identity information from low-quality images and then enhances texture details according to the distribution of real faces. This design is implemented with two key components: 1) Identity Restoration Module (IRM) for preserving the face details in results. Instead of denoising from pure Gaussian random distribution with LQ images as the condition during the reverse process, we propose a novel truncated sampling method which starts from LQ images with part noise added. We theoretically prove that this change shrinks the evidence lower bound of DPM and then restores more original details. With theoretical proof, two cascade conditional DPMs with different input sizes are introduced to strengthen this sampling effect and reduce training difficulty in the high-resolution image generated directly. 2) Texture Enhancement Module (TEM) for polishing the texture of the image. Here an unconditional DPM, a LQ-free model, is introduced to further force the restorations to appear realistic. We theoretically proved that this unconditional DPM trained on pure HQ images contributes to justifying the correct distribution of inference images output from IRM in pixel-level space. Truncated sampling with fractional time step is utilized to polish pixel-level textures while preserving identity information.