Pose and body shape editing in a human image has received increasing attention. However, current methods often struggle with dataset biases and deteriorate realism and the person's identity when users make large edits. We propose a one-shot approach that enables large edits with identity preservation. To enable large edits, we fit a 3D body model, project the input image onto the 3D model, and change the body's pose and shape. Because this initial textured body model has artifacts due to occlusion and the inaccurate body shape, the rendered image undergoes a diffusion-based refinement, in which strong noise destroys body structure and identity whereas insufficient noise does not help. We thus propose an iterative refinement with weak noise, applied first for the whole body and then for the face. We further enhance the realism by fine-tuning text embeddings via self-supervised learning. Our quantitative and qualitative evaluations demonstrate that our method outperforms other existing methods across various datasets.
Spiking Neural Networks (SNNs), known for their biologically plausible architecture, face the challenge of limited performance. The self-attention mechanism, which is the cornerstone of the high-performance Transformer and also a biologically inspired structure, is absent in existing SNNs. To this end, we explore the potential of leveraging both self-attention capability and biological properties of SNNs, and propose a novel Spiking Self-Attention (SSA) and Spiking Transformer (Spikformer). The SSA mechanism eliminates the need for softmax and captures the sparse visual feature employing spike-based Query, Key, and Value. This sparse computation without multiplication makes SSA efficient and energy-saving. Further, we develop a Spiking Convolutional Stem (SCS) with supplementary convolutional layers to enhance the architecture of Spikformer. The Spikformer enhanced with the SCS is referred to as Spikformer V2. To train larger and deeper Spikformer V2, we introduce a pioneering exploration of Self-Supervised Learning (SSL) within the SNN. Specifically, we pre-train Spikformer V2 with masking and reconstruction style inspired by the mainstream self-supervised Transformer, and then finetune the Spikformer V2 on the image classification on ImageNet. Extensive experiments show that Spikformer V2 outperforms other previous surrogate training and ANN2SNN methods. An 8-layer Spikformer V2 achieves an accuracy of 80.38% using 4 time steps, and after SSL, a 172M 16-layer Spikformer V2 reaches an accuracy of 81.10% with just 1 time step. To the best of our knowledge, this is the first time that the SNN achieves 80+% accuracy on ImageNet. The code will be available at Spikformer V2.
We propose Guided Positive Sampling Self-Supervised Learning (GPS-SSL), a general method to inject a priori knowledge into Self-Supervised Learning (SSL) positive samples selection. Current SSL methods leverage Data-Augmentations (DA) for generating positive samples and incorporate prior knowledge - an incorrect, or too weak DA will drastically reduce the quality of the learned representation. GPS-SSL proposes instead to design a metric space where Euclidean distances become a meaningful proxy for semantic relationship. In that space, it is now possible to generate positive samples from nearest neighbor sampling. Any prior knowledge can now be embedded into that metric space independently from the employed DA. From its simplicity, GPS-SSL is applicable to any SSL method, e.g. SimCLR or BYOL. A key benefit of GPS-SSL is in reducing the pressure in tailoring strong DAs. For example GPS-SSL reaches 85.58% on Cifar10 with weak DA while the baseline only reaches 37.51%. We therefore move a step forward towards the goal of making SSL less reliant on DA. We also show that even when using strong DAs, GPS-SSL outperforms the baselines on under-studied domains. We evaluate GPS-SSL along with multiple baseline SSL methods on numerous downstream datasets from different domains when the models use strong or minimal data augmentations. We hope that GPS-SSL will open new avenues in studying how to inject a priori knowledge into SSL in a principled manner.
Self-supervised pre-training paradigms have been extensively explored in the field of skeleton-based action recognition. In particular, methods based on masked prediction have pushed the performance of pre-training to a new height. However, these methods take low-level features, such as raw joint coordinates or temporal motion, as prediction targets for the masked regions, which is suboptimal. In this paper, we show that using high-level contextualized features as prediction targets can achieve superior performance. Specifically, we propose Skeleton2vec, a simple and efficient self-supervised 3D action representation learning framework, which utilizes a transformer-based teacher encoder taking unmasked training samples as input to create latent contextualized representations as prediction targets. Benefiting from the self-attention mechanism, the latent representations generated by the teacher encoder can incorporate the global context of the entire training samples, leading to a richer training task. Additionally, considering the high temporal correlations in skeleton sequences, we propose a motion-aware tube masking strategy which divides the skeleton sequence into several tubes and performs persistent masking within each tube based on motion priors, thus forcing the model to build long-range spatio-temporal connections and focus on action-semantic richer regions. Extensive experiments on NTU-60, NTU-120, and PKU-MMD datasets demonstrate that our proposed Skeleton2vec outperforms previous methods and achieves state-of-the-art results.
Early cancer detection is crucial for prognosis, but many cancer types lack large labelled datasets required for developing deep learning models. This paper investigates self-supervised learning (SSL) as an alternative to the standard supervised pre-training on ImageNet for scenarios with limited training data using a deep learning model (ResNet-50). We first demonstrate that SSL pre-training on ImageNet (via the Barlow Twins SSL algorithm) outperforms supervised pre-training (SL) using a skin lesion dataset with limited training samples. We then consider \textit{further} SSL pre-training (of the two ImageNet pre-trained models) on task-specific datasets, where our implementation is motivated by supervised transfer learning. This approach significantly enhances initially SL pre-trained models, closing the performance gap with initially SSL pre-trained ones. Surprisingly, further pre-training on just the limited fine-tuning data achieves this performance equivalence. Linear probe experiments reveal that improvement stems from enhanced feature extraction. Hence, we find that minimal further SSL pre-training on task-specific data can be as effective as large-scale SSL pre-training on ImageNet for medical image classification tasks with limited labelled data. We validate these results on an oral cancer histopathology dataset, suggesting broader applicability across medical imaging domains facing labelled data scarcity.
Video-based facial affect analysis has recently attracted increasing attention owing to its critical role in human-computer interaction. Previous studies mainly focus on developing various deep learning architectures and training them in a fully supervised manner. Although significant progress has been achieved by these supervised methods, the longstanding lack of large-scale high-quality labeled data severely hinders their further improvements. Motivated by the recent success of self-supervised learning in computer vision, this paper introduces a self-supervised approach, termed Self-supervised Video Facial Affect Perceiver (SVFAP), to address the dilemma faced by supervised methods. Specifically, SVFAP leverages masked facial video autoencoding to perform self-supervised pre-training on massive unlabeled facial videos. Considering that large spatiotemporal redundancy exists in facial videos, we propose a novel temporal pyramid and spatial bottleneck Transformer as the encoder of SVFAP, which not only largely reduces computational costs but also achieves excellent performance. To verify the effectiveness of our method, we conduct experiments on nine datasets spanning three downstream tasks, including dynamic facial expression recognition, dimensional emotion recognition, and personality recognition. Comprehensive results demonstrate that SVFAP can learn powerful affect-related representations via large-scale self-supervised pre-training and it significantly outperforms previous state-of-the-art methods on all datasets. Code is available at https://github.com/sunlicai/SVFAP.
Masked image modeling (MIM) has emerged as a promising approach for training Vision Transformers (ViTs). The essence of MIM lies in the token-wise prediction of masked tokens, which aims to predict targets tokenized from images or generated by pre-trained models like vision-language models. While using tokenizers or pre-trained models are plausible MIM targets, they often offer spatially inconsistent targets even for neighboring tokens, complicating models to learn unified and discriminative representations. Our pilot study identifies spatial inconsistencies and suggests that resolving them can accelerate representation learning. Building upon this insight, we introduce a novel self-supervision signal called Dynamic Token Morphing (DTM), which dynamically aggregates contextually related tokens to yield contextualized targets, thereby mitigating spatial inconsistency. DTM is compatible with various SSL frameworks; we showcase improved MIM results by employing DTM, barely introducing extra training costs. Our method facilitates training by using consistent targets, resulting in 1) faster training and 2) reduced losses. Experiments on ImageNet-1K and ADE20K demonstrate the superiority of our method compared with state-of-the-art, complex MIM methods. Furthermore, the comparative evaluation of the iNaturalists and fine-grained visual classification datasets further validates the transferability of our method on various downstream tasks. Code is available at https://github.com/naver-ai/dtm
As the deep learning revolution marches on, self-supervised learning has garnered increasing attention in recent years thanks to its remarkable representation learning ability and the low dependence on labeled data. Among these varied self-supervised techniques, masked modeling has emerged as a distinctive approach that involves predicting parts of the original data that are proportionally masked during training. This paradigm enables deep models to learn robust representations and has demonstrated exceptional performance in the context of computer vision, natural language processing, and other modalities. In this survey, we present a comprehensive review of the masked modeling framework and its methodology. We elaborate on the details of techniques within masked modeling, including diverse masking strategies, recovering targets, network architectures, and more. Then, we systematically investigate its wide-ranging applications across domains. Furthermore, we also explore the commonalities and differences between masked modeling methods in different fields. Toward the end of this paper, we conclude by discussing the limitations of current techniques and point out several potential avenues for advancing masked modeling research. A paper list project with this survey is available at \url{https://github.com/Lupin1998/Awesome-MIM}.
The capability to accurately estimate 3D human poses is crucial for diverse fields such as action recognition, gait recognition, and virtual/augmented reality. However, a persistent and significant challenge within this field is the accurate prediction of human poses under conditions of severe occlusion. Traditional image-based estimators struggle with heavy occlusions due to a lack of temporal context, resulting in inconsistent predictions. While video-based models benefit from processing temporal data, they encounter limitations when faced with prolonged occlusions that extend over multiple frames. This challenge arises because these models struggle to generalize beyond their training datasets, and the variety of occlusions is hard to capture in the training data. Addressing these challenges, we propose STRIDE (Single-video based TempoRally contInuous Occlusion-Robust 3D Pose Estimation), a novel Test-Time Training (TTT) approach to fit a human motion prior for each video. This approach specifically handles occlusions that were not encountered during the model's training. By employing STRIDE, we can refine a sequence of noisy initial pose estimates into accurate, temporally coherent poses during test time, effectively overcoming the limitations of prior methods. Our framework demonstrates flexibility by being model-agnostic, allowing us to use any off-the-shelf 3D pose estimation method for improving robustness and temporal consistency. We validate STRIDE's efficacy through comprehensive experiments on challenging datasets like Occluded Human3.6M, Human3.6M, and OCMotion, where it not only outperforms existing single-image and video-based pose estimation models but also showcases superior handling of substantial occlusions, achieving fast, robust, accurate, and temporally consistent 3D pose estimates. Code is made publicly available at https://github.com/take2rohit/stride
This study utilizes community structures to address node degree biases in message-passing (MP) via learnable graph augmentations and novel graph transformers. Recent augmentation-based methods showed that MP neural networks often perform poorly on low-degree nodes, leading to degree biases due to a lack of messages reaching low-degree nodes. Despite their success, most methods use heuristic or uniform random augmentations, which are non-differentiable and may not always generate valuable edges for learning representations. In this paper, we propose Community-aware Graph Transformers, namely CGT, to learn degree-unbiased representations based on learnable augmentations and graph transformers by extracting within community structures. We first design a learnable graph augmentation to generate more within-community edges connecting low-degree nodes through edge perturbation. Second, we propose an improved self-attention to learn underlying proximity and the roles of nodes within the community. Third, we propose a self-supervised learning task that could learn the representations to preserve the global graph structure and regularize the graph augmentations. Extensive experiments on various benchmark datasets showed CGT outperforms state-of-the-art baselines and significantly improves the node degree biases. The source code is available at https://github.com/NSLab-CUK/Community-aware-Graph-Transformer.
Self-supervised learning (SSL) in audio holds significant potential across various domains, particularly in situations where abundant, unlabeled data is readily available at no cost. This is pertinent in bioacoustics, where biologists routinely collect extensive sound datasets from the natural environment. In this study, we demonstrate that SSL is capable of acquiring meaningful representations of bird sounds from audio recordings without the need for annotations. Our experiments showcase that these learned representations exhibit the capacity to generalize to new bird species in few-shot learning (FSL) scenarios. Additionally, we show that selecting windows with high bird activation for self-supervised learning, using a pretrained audio neural network, significantly enhances the quality of the learned representations.
Contrastive representation learning has emerged as an outstanding approach for anomaly detection. In this work, we explore the $\ell_2$-norm of contrastive features and its applications in out-of-distribution detection. We propose a simple method based on contrastive learning, which incorporates out-of-distribution data by discriminating against normal samples in the contrastive layer space. Our approach can be applied flexibly as an outlier exposure (OE) approach, where the out-of-distribution data is a huge collective of random images, or as a fully self-supervised learning approach, where the out-of-distribution data is self-generated by applying distribution-shifting transformations. The ability to incorporate additional out-of-distribution samples enables a feasible solution for datasets where AD methods based on contrastive learning generally underperform, such as aerial images or microscopy images. Furthermore, the high-quality features learned through contrastive learning consistently enhance performance in OE scenarios, even when the available out-of-distribution dataset is not diverse enough. Our extensive experiments demonstrate the superiority of our proposed method under various scenarios, including unimodal and multimodal settings, with various image datasets.
The objective of Active Learning is to strategically label a subset of the dataset to maximize performance within a predetermined labeling budget. In this study, we harness features acquired through self-supervised learning. We introduce a straightforward yet potent metric, Cluster Distance Difference, to identify diverse data. Subsequently, we introduce a novel framework, Balancing Active Learning (BAL), which constructs adaptive sub-pools to balance diverse and uncertain data. Our approach outperforms all established active learning methods on widely recognized benchmarks by 1.20%. Moreover, we assess the efficacy of our proposed framework under extended settings, encompassing both larger and smaller labeling budgets. Experimental results demonstrate that, when labeling 80% of the samples, the performance of the current SOTA method declines by 0.74%, whereas our proposed BAL achieves performance comparable to the full dataset. Codes are available at https://github.com/JulietLJY/BAL.
We propose emotion2vec, a universal speech emotion representation model. emotion2vec is pre-trained on open-source unlabeled emotion data through self-supervised online distillation, combining utterance-level loss and frame-level loss during pre-training. emotion2vec outperforms state-of-the-art pre-trained universal models and emotion specialist models by only training linear layers for the speech emotion recognition task on the mainstream IEMOCAP dataset. In addition, emotion2vec shows consistent improvements among 10 different languages of speech emotion recognition datasets. emotion2vec also shows excellent results on other emotion tasks, such as song emotion recognition, emotion prediction in conversation, and sentiment analysis. Comparison experiments, ablation experiments, and visualization comprehensively demonstrate the universal capability of the proposed emotion2vec. To the best of our knowledge, emotion2vec is the first universal representation model in various emotion-related tasks, filling a gap in the field.
Recently, model-based reinforcement learning algorithms have demonstrated remarkable efficacy in visual input environments. These approaches begin by constructing a parameterized simulation world model of the real environment through self-supervised learning. By leveraging the imagination of the world model, the agent's policy is enhanced without the constraints of sampling from the real environment. The performance of these algorithms heavily relies on the sequence modeling and generation capabilities of the world model. However, constructing a perfectly accurate model of a complex unknown environment is nearly impossible. Discrepancies between the model and reality may cause the agent to pursue virtual goals, resulting in subpar performance in the real environment. Introducing random noise into model-based reinforcement learning has been proven beneficial. In this work, we introduce Stochastic Transformer-based wORld Model (STORM), an efficient world model architecture that combines the strong sequence modeling and generation capabilities of Transformers with the stochastic nature of variational autoencoders. STORM achieves a mean human performance of $126.7\%$ on the Atari $100$k benchmark, setting a new record among state-of-the-art methods that do not employ lookahead search techniques. Moreover, training an agent with $1.85$ hours of real-time interaction experience on a single NVIDIA GeForce RTX 3090 graphics card requires only $4.3$ hours, showcasing improved efficiency compared to previous methodologies.
Conventional audio-visual approaches for active speaker detection (ASD) typically rely on visually pre-extracted face tracks and the corresponding single-channel audio to find the speaker in a video. Therefore, they tend to fail every time the face of the speaker is not visible. We demonstrate that a simple audio convolutional recurrent neural network (CRNN) trained with spatial input features extracted from multichannel audio can perform simultaneous horizontal active speaker detection and localization (ASDL), independently of the visual modality. To address the time and cost of generating ground truth labels to train such a system, we propose a new self-supervised training pipeline that embraces a ``student-teacher'' learning approach. A conventional pre-trained active speaker detector is adopted as a ``teacher'' network to provide the position of the speakers as pseudo-labels. The multichannel audio ``student'' network is trained to generate the same results. At inference, the student network can generalize and locate also the occluded speakers that the teacher network is not able to detect visually, yielding considerable improvements in recall rate. Experiments on the TragicTalkers dataset show that an audio network trained with the proposed self-supervised learning approach can exceed the performance of the typical audio-visual methods and produce results competitive with the costly conventional supervised training. We demonstrate that improvements can be achieved when minimal manual supervision is introduced in the learning pipeline. Further gains may be sought with larger training sets and integrating vision with the multichannel audio system.
Motivated by high resource costs of centralized machine learning schemes as well as data privacy concerns, federated learning (FL) emerged as an efficient alternative that relies on aggregating locally trained models rather than collecting clients' potentially private data. In practice, available resources and data distributions vary from one client to another, creating an inherent system heterogeneity that leads to deterioration of the performance of conventional FL algorithms. In this work, we present a federated quantization-based self-supervised learning scheme (Fed-QSSL) designed to address heterogeneity in FL systems. At clients' side, to tackle data heterogeneity we leverage distributed self-supervised learning while utilizing low-bit quantization to satisfy constraints imposed by local infrastructure and limited communication resources. At server's side, Fed-QSSL deploys de-quantization, weighted aggregation and re-quantization, ultimately creating models personalized to both data distribution as well as specific infrastructure of each client's device. We validated the proposed algorithm on real world datasets, demonstrating its efficacy, and theoretically analyzed impact of low-bit training on the convergence and robustness of the learned models.
Continued pre-training (CP) offers multiple advantages, like target domain adaptation and the potential to exploit the continuous stream of unlabeled data available online. However, continued pre-training on out-of-domain distributions often leads to catastrophic forgetting of previously acquired knowledge, leading to sub-optimal ASR performance. This paper presents FusDom, a simple and novel methodology for SSL-based continued pre-training. FusDom learns speech representations that are robust and adaptive yet not forgetful of concepts seen in the past. Instead of solving the SSL pre-text task on the output representations of a single model, FusDom leverages two identical pre-trained SSL models, a teacher and a student, with a modified pre-training head to solve the CP SSL pre-text task. This head employs a cross-attention mechanism between the representations of both models while only the student receives gradient updates and the teacher does not. Finally, the student is fine-tuned for ASR. In practice, FusDom outperforms all our baselines across settings significantly, with WER improvements in the range of 0.2 WER - 7.3 WER in the target domain while retaining the performance in the earlier domain.
Self-supervised methods have recently proved to be nearly as effective as supervised ones in various imaging inverse problems, paving the way for learning-based approaches in scientific and medical imaging applications where ground truth data is hard or expensive to obtain. These methods critically rely on invariance to translations and/or rotations of the image distribution to learn from incomplete measurement data alone. However, existing approaches fail to obtain competitive performances in the problems of image super-resolution and deblurring, which play a key role in most imaging systems. In this work, we show that invariance to roto-translations is insufficient to learn from measurements that only contain low-frequency information. Instead, we propose scale-equivariant imaging, a new self-supervised approach that leverages the fact that many image distributions are approximately scale-invariant, enabling the recovery of high-frequency information lost in the measurement process. We demonstrate throughout a series of experiments on real datasets that the proposed method outperforms other self-supervised approaches, and obtains performances on par with fully supervised learning.
The scarcity of annotated data in LiDAR point cloud understanding hinders effective representation learning. Consequently, scholars have been actively investigating efficacious self-supervised pre-training paradigms. Nevertheless, temporal information, which is inherent in the LiDAR point cloud sequence, is consistently disregarded. To better utilize this property, we propose an effective pre-training strategy, namely Temporal Masked Auto-Encoders (T-MAE), which takes as input temporally adjacent frames and learns temporal dependency. A SiamWCA backbone, containing a Siamese encoder and a windowed cross-attention (WCA) module, is established for the two-frame input. Considering that the movement of an ego-vehicle alters the view of the same instance, temporal modeling also serves as a robust and natural data augmentation, enhancing the comprehension of target objects. SiamWCA is a powerful architecture but heavily relies on annotated data. Our T-MAE pre-training strategy alleviates its demand for annotated data. Comprehensive experiments demonstrate that T-MAE achieves the best performance on both Waymo and ONCE datasets among competitive self-supervised approaches. Codes will be released at https://github.com/codename1995/T-MAE