Recent work in Video Frame Interpolation (VFI) tries to formulate VFI as a diffusion-based conditional image generation problem, synthesizing the intermediate frame given a random noise and neighboring frames. Due to the relatively high resolution of videos, Latent Diffusion Models (LDMs) are employed as the conditional generation model, where the autoencoder compresses images into latent representations for diffusion and then reconstructs images from these latent representations. Such a formulation poses a crucial challenge: VFI expects that the output is deterministically equal to the ground truth intermediate frame, but LDMs randomly generate a diverse set of different images when the model runs multiple times. The reason for the diverse generation is that the cumulative variance (variance accumulated at each step of generation) of generated latent representations in LDMs is large. This makes the sampling trajectory random, resulting in diverse rather than deterministic generations. To address this problem, we propose our unique solution: Frame Interpolation with Consecutive Brownian Bridge Diffusion. Specifically, we propose consecutive Brownian Bridge diffusion that takes a deterministic initial value as input, resulting in a much smaller cumulative variance of generated latent representations. Our experiments suggest that our method can improve together with the improvement of the autoencoder and achieve state-of-the-art performance in VFI, leaving strong potential for further enhancement.
Conventional video object segmentation (VOS) methods usually necessitate a substantial volume of pixel-level annotated video data for fully supervised learning. In this paper, we present HVC, a \textbf{h}ybrid static-dynamic \textbf{v}isual \textbf{c}orrespondence framework for self-supervised VOS. HVC extracts pseudo-dynamic signals from static images, enabling an efficient and scalable VOS model. Our approach utilizes a minimalist fully-convolutional architecture to capture static-dynamic visual correspondence in image-cropped views. To achieve this objective, we present a unified self-supervised approach to learn visual representations of static-dynamic feature similarity. Firstly, we establish static correspondence by utilizing a priori coordinate information between cropped views to guide the formation of consistent static feature representations. Subsequently, we devise a concise convolutional layer to capture the forward / backward pseudo-dynamic signals between two views, serving as cues for dynamic representations. Finally, we propose a hybrid visual correspondence loss to learn joint static and dynamic consistency representations. Our approach, without bells and whistles, necessitates only one training session using static image data, significantly reducing memory consumption ($\sim$16GB) and training time ($\sim$\textbf{2h}). Moreover, HVC achieves state-of-the-art performance in several self-supervised VOS benchmarks and additional video label propagation tasks.
Human perception of the world is shaped by a multitude of viewpoints and modalities. While many existing datasets focus on scene understanding from a certain perspective (e.g. egocentric or third-person views), our dataset offers a panoptic perspective (i.e. multiple viewpoints with multiple data modalities). Specifically, we encapsulate third-person panoramic and front views, as well as egocentric monocular/binocular views with rich modalities including video, multi-channel audio, directional binaural delay, location data and textual scene descriptions within each scene captured, presenting comprehensive observation of the world. Figure 1 offers a glimpse of all 28 scene categories of our 360+x dataset. To the best of our knowledge, this is the first database that covers multiple viewpoints with multiple data modalities to mimic how daily information is accessed in the real world. Through our benchmark analysis, we presented 5 different scene understanding tasks on the proposed 360+x dataset to evaluate the impact and benefit of each data modality and perspective in panoptic scene understanding. We hope this unique dataset could broaden the scope of comprehensive scene understanding and encourage the community to approach these problems from more diverse perspectives.
High-quality human reconstruction and photo-realistic rendering of a dynamic scene is a long-standing problem in computer vision and graphics. Despite considerable efforts invested in developing various capture systems and reconstruction algorithms, recent advancements still struggle with loose or oversized clothing and overly complex poses. In part, this is due to the challenges of acquiring high-quality human datasets. To facilitate the development of these fields, in this paper, we present PKU-DyMVHumans, a versatile human-centric dataset for high-fidelity reconstruction and rendering of dynamic human scenarios from dense multi-view videos. It comprises 8.2 million frames captured by more than 56 synchronized cameras across diverse scenarios. These sequences comprise 32 human subjects across 45 different scenarios, each with a high-detailed appearance and realistic human motion. Inspired by recent advancements in neural radiance field (NeRF)-based scene representations, we carefully set up an off-the-shelf framework that is easy to provide those state-of-the-art NeRF-based implementations and benchmark on PKU-DyMVHumans dataset. It is paving the way for various applications like fine-grained foreground/background decomposition, high-quality human reconstruction and photo-realistic novel view synthesis of a dynamic scene. Extensive studies are performed on the benchmark, demonstrating new observations and challenges that emerge from using such high-fidelity dynamic data.
Auditory and visual signals usually present together and correlate with each other, not only in natural environments but also in clinical settings. However, the audio-visual modelling in the latter case can be more challenging, due to the different sources of audio/video signals and the noise (both signal-level and semantic-level) in auditory signals -- usually speech. In this paper, we consider audio-visual modelling in a clinical setting, providing a solution to learn medical representations that benefit various clinical tasks, without human expert annotation. A simple yet effective multi-modal self-supervised learning framework is proposed for this purpose. The proposed approach is able to localise anatomical regions of interest during ultrasound imaging, with only speech audio as a reference. Experimental evaluations on a large-scale clinical multi-modal ultrasound video dataset show that the proposed self-supervised method learns good transferable anatomical representations that boost the performance of automated downstream clinical tasks, even outperforming fully-supervised solutions.
Class incremental semantic segmentation aims to strike a balance between the model's stability and plasticity by maintaining old knowledge while adapting to new concepts. However, most state-of-the-art methods use the freeze strategy for stability, which compromises the model's plasticity.In contrast, releasing parameter training for plasticity could lead to the best performance for all categories, but this requires discriminative feature representation.Therefore, we prioritize the model's plasticity and propose the Contrast inter- and intra-class representations for Incremental Segmentation (CoinSeg), which pursues discriminative representations for flexible parameter tuning. Inspired by the Gaussian mixture model that samples from a mixture of Gaussian distributions, CoinSeg emphasizes intra-class diversity with multiple contrastive representation centroids. Specifically, we use mask proposals to identify regions with strong objectness that are likely to be diverse instances/centroids of a category. These mask proposals are then used for contrastive representations to reinforce intra-class diversity. Meanwhile, to avoid bias from intra-class diversity, we also apply category-level pseudo-labels to enhance category-level consistency and inter-category diversity. Additionally, CoinSeg ensures the model's stability and alleviates forgetting through a specific flexible tuning strategy. We validate CoinSeg on Pascal VOC 2012 and ADE20K datasets with multiple incremental scenarios and achieve superior results compared to previous state-of-the-art methods, especially in more challenging and realistic long-term scenarios. Code is available at https://github.com/zkzhang98/CoinSeg.
This work aims to improve unsupervised audio-visual pre-training. Inspired by the efficacy of data augmentation in visual contrastive learning, we propose a novel speed co-augmentation method that randomly changes the playback speeds of both audio and video data. Despite its simplicity, the speed co-augmentation method possesses two compelling attributes: (1) it increases the diversity of audio-visual pairs and doubles the size of negative pairs, resulting in a significant enhancement in the learned representations, and (2) it changes the strict correlation between audio-visual pairs but introduces a partial relationship between the augmented pairs, which is modeled by our proposed SoftInfoNCE loss to further boost the performance. Experimental results show that the proposed method significantly improves the learned representations when compared to vanilla audio-visual contrastive learning.
With its significant performance improvements, the deep learning paradigm has become a standard tool for modern image denoisers. While promising performance has been shown on seen noise distributions, existing approaches often suffer from generalisation to unseen noise types or general and real noise. It is understandable as the model is designed to learn paired mapping (e.g. from a noisy image to its clean version). In this paper, we instead propose to learn to disentangle the noisy image, under the intuitive assumption that different corrupted versions of the same clean image share a common latent space. A self-supervised learning framework is proposed to achieve the goal, without looking at the latent clean image. By taking two different corrupted versions of the same image as input, the proposed Multi-view Self-supervised Disentanglement (MeD) approach learns to disentangle the latent clean features from the corruptions and recover the clean image consequently. Extensive experimental analysis on both synthetic and real noise shows the superiority of the proposed method over prior self-supervised approaches, especially on unseen novel noise types. On real noise, the proposed method even outperforms its supervised counterparts by over 3 dB.
Deep learning based medical volumetric segmentation methods either train the model from scratch or follow the standard "pre-training then finetuning" paradigm. Although finetuning a well pre-trained model on downstream tasks can harness its representation power, the standard full finetuning is costly in terms of computation and memory footprint. In this paper, we present the first study on parameter-efficient transfer learning for medical volumetric segmentation and propose a novel framework named Med-Tuning based on intra-stage feature enhancement and inter-stage feature interaction. Given a large-scale pre-trained model on 2D natural images, our method can exploit both the multi-scale spatial feature representations and temporal correlations along image slices, which are crucial for accurate medical volumetric segmentation. Extensive experiments on three benchmark datasets (including CT and MRI) show that our method can achieve better results than previous state-of-the-art parameter-efficient transfer learning methods and full finetuning for the segmentation task, with much less tuned parameter costs. Compared to full finetuning, our method reduces the finetuned model parameters by up to 4x, with even better segmentation performance.
The research community has witnessed the powerful potential of self-supervised Masked Image Modeling (MIM), which enables the models capable of learning visual representation from unlabeled data. In this paper, to incorporate both the crucial global structural information and local details for dense prediction tasks, we alter the perspective to the frequency domain and present a new MIM-based framework named FreMAE for self-supervised pre-training for medical image segmentation. Based on the observations that the detailed structural information mainly lies in the high-frequency components and the high-level semantics are abundant in the low-frequency counterparts, we further incorporate multi-stage supervision to guide the representation learning during the pre-training phase. Extensive experiments on three benchmark datasets show the superior advantage of our proposed FreMAE over previous state-of-the-art MIM methods. Compared with various baselines trained from scratch, our FreMAE could consistently bring considerable improvements to the model performance. To the best our knowledge, this is the first attempt towards MIM with Fourier Transform in medical image segmentation.