Masked Image Modeling (MIM) has emerged as a powerful self-supervised learning paradigm for visual representation learning, enabling models to acquire rich visual representations by predicting masked portions of images from their visible regions. While this approach has shown promising results, we hypothesize that its effectiveness may be limited by optimization challenges during early training stages, where models are expected to learn complex image distributions from partial observations before developing basic visual processing capabilities. To address this limitation, we propose a prototype-driven curriculum learning framework that structures the learning process to progress from prototypical examples to more complex variations in the dataset. Our approach introduces a temperature-based annealing scheme that gradually expands the training distribution, enabling more stable and efficient learning trajectories. Through extensive experiments on ImageNet-1K, we demonstrate that our curriculum learning strategy significantly improves both training efficiency and representation quality while requiring substantially fewer training epochs compared to standard Masked Auto-Encoding. Our findings suggest that carefully controlling the order of training examples plays a crucial role in self-supervised visual learning, providing a practical solution to the early-stage optimization challenges in MIM.

Contrastive image-text pre-trained models such as CLIP have shown remarkable adaptability to downstream tasks. However, they face challenges due to the high computational requirements of the Vision Transformer (ViT) backbone. Current strategies to boost ViT efficiency focus on pruning patch tokens but fall short in addressing the multimodal nature of CLIP and identifying the optimal subset of tokens for maximum performance. To address this, we propose greedy search methods to establish a “Golden Ranking” and introduce a lightweight predictor specifically trained to approximate this Ranking. To compensate for any performance degradation resulting from token pruning, we incorporate learnable visual tokens that aid in restoring and potentially enhancing the model's performance. Our work presents a comprehensive and systematic investigation of pruning tokens within the ViT backbone of CLIP models. Through our framework, we successfully reduced 40% of patch tokens in CLIP's ViT while only suffering a minimal average accuracy loss of 0.3% across seven datasets. Our study lays the groundwork for building more computationally efficient multimodal models without sacrificing their performance, addressing a key challenge in the application of advanced vision-language models.11Project Page: https://github.com/CEWu/PatchRanking

Humans possess a remarkable ability to integrate auditory and visual information, enabling a deeper understanding of the surrounding environment. This early fusion of audio and visual cues, demonstrated through cognitive psychology and neuroscience research, offers promising potential for developing multimodal perception models. However, training early fusion architectures poses significant challenges, as the increased model expressivity requires robust learning frameworks to harness their enhanced capabilities. In this paper, we address this challenge by leveraging the masked reconstruction framework, previously successful in unimodal settings, to train audio-visual encoders with early fusion. Additionally, we propose an attention-based fusion module that captures interactions between local audio and visual representations, enhancing the model's ability to capture fine-grained interactions. While effective, this procedure can become computationally intractable, as the number of local representations increases. Thus, to address the computational complexity, we propose an alternative procedure that factorizes the local representations before representing audio-visual interactions. Extensive evaluations on a variety of datasets demonstrate the superiority of our approach in audio-event classification, visual sound localization, sound separation, and audio-visual segmentation. These contributions enable the efficient training of deeply integrated audio-visual models and significantly advance the usefulness of early fusion architectures.

Our work tackles the computational challenges of contrastive learning methods, particularly for the pretraining of Vision Transformers (ViTs). Despite the effectiveness of contrastive learning, the substantial computational resources required for training often hinder their practical application. To mitigate this issue, we propose an acceleration framework, leveraging ViT's unique ability to generalize across inputs of varying sequence lengths. Our method employs a mix of sequence compression strategies, including randomized token dropout and flexible patch scaling, to reduce the cost of gradient estimation and accelerate convergence. We further provide an in-depth analysis of the gradient estimation error of various acceleration strategies as well as their impact on downstream tasks, offering valuable insights into the trade-offs between acceleration and performance. We also propose a novel procedure to identify an optimal acceleration schedule to adjust the sequence compression ratios to the training progress, ensuring efficient training without sacrificing downstream performance. Our approach significantly reduces computational overhead across various self-supervised learning algorithms on large-scale datasets. In ImageNet, our method achieves speedups of 4$\times$ in MoCo, 3.3$\times$ in SimCLR, and 2.5$\times$ in DINO, demonstrating substantial efficiency gains.

Vision-language models such as CLIP [27] learn a generic text-image embedding from large-scale training data. A vision-language model can be adapted to a new classification task through few-shot prompt tuning. We find that such a prompt tuning process is highly robust to label noises. This intrigues us to study the key reasons contributing to the robustness of the prompt tuning paradigm. We conducted extensive experiments to explore this property and find the key factors are: 1) the fixed classname tokens provide a strong regularization to the optimization of the model, reducing gradients induced by the noisy samples; 2) the powerful pre-trained image-text embedding that is learned from diverse and generic web data provides strong prior knowledge for image classification. Further, we demonstrate that noisy zero-shot predictions from CLIP can be used to tune its own prompt, significantly enhancing prediction accuracy in the unsupervised setting. The code is available at https://github.com/CEWu/PTNL.

The ability to accurately recognize, localize and separate sound sources is fundamental to any audio-visual perception task. Historically, these abilities were tackled separately, with several methods developed independently for each task. However, given the interconnected nature of source localization, separation, and recognition, independent models are likely to yield suboptimal performance as they fail to capture the interdependence between these tasks. To address this problem, we propose a unified audio-visual learning framework (dubbed OneAVM) that integrates audio and visual cues for joint localization, separation, and recognition. OneAVM comprises a shared audio-visual encoder and task-specific decoders trained with three objectives. The first objective aligns audio and visual representations through a localized audio-visual correspondence loss. The second tackles visual source separation using a traditional mix-and-separate framework. Finally, the third objective reinforces visual feature separation and localization by mixing images in pixel space and aligning their representations with those of all corresponding sound sources. Extensive experiments on MUSIC, VGG-Instruments, VGG-Music, and VGGSound datasets demonstrate the effectiveness of OneAVM for all three tasks, audio-visual source localization, separation, and nearest neighbor recognition, and empirically demonstrate a strong positive transfer between them.

We propose a self-supervised algorithm to learn representations from egocentric video data. Recently, significant efforts have been made to capture humans interacting with their own environments as they go about their daily activities. In result, several large egocentric datasets of interaction-rich multi-modal data have emerged. However, learning representations from videos can be challenging. First, given the uncurated nature of long-form continuous videos, learning effective representations require focusing on moments in time when interactions take place. Second, visual representations of daily activities should be sensitive to changes in the state of the environment. However, current successful multi-modal learning frameworks encourage representation invariance over time. To address these challenges, we leverage audio signals to identify moments of likely interactions which are conducive to better learning. We also propose a novel self-supervised objective that learns from audible state changes caused by interactions. We validate these contributions extensively on two large-scale egocentric datasets, EPIC-Kitchens-100 and the recently released Ego4D, and show improvements on several downstream tasks, including action recognition, long-term action anticipation, and object state change classification.

Audio-visual source localization is a challenging task that aims to predict the location of visual sound sources in a video. Since collecting ground-truth annotations of sounding objects can be costly, a plethora of weakly-supervised localization methods that can learn from datasets with no bounding-box annotations have been proposed in recent years, by leveraging the natural co-occurrence of audio and visual signals. Despite significant interest, popular evaluation protocols have two major flaws. First, they allow for the use of a fully annotated dataset to perform early stopping, thus significantly increasing the annotation effort required for training. Second, current evaluation metrics assume the presence of sound sources at all times. This is of course an unrealistic assumption, and thus better metrics are necessary to capture the model's performance on (negative) samples with no visible sound sources. To accomplish this, we extend the test set of popular benchmarks, Flickr SoundNet and VGG-Sound Sources, in order to include negative samples, and measure performance using metrics that balance localization accuracy and recall. Using the new protocol, we conducted an extensive evaluation of prior methods, and found that most prior works are not capable of identifying negatives and suffer from significant overfitting problems (rely heavily on early stopping for best results). We also propose a new approach for visual sound source localization that addresses both these problems. In particular, we found that, through extreme visual dropout and the use of momentum encoders, the proposed approach combats overfitting effectively, and establishes a new state-of-the-art performance on both Flickr SoundNet and VGG-Sound Source. Code and pre-trained models are available at https://github.com/stoneMo/SLAVC.

We present a self-supervised learning method to learn audio and video representations. Prior work uses the natural correspondence between audio and video to define a standard cross-modal instance discrimination task, where a model is trained to match representations from the two modalities. However, the standard approach introduces two sources of training noise. First, audio-visual correspondences often produce faulty positives since the audio and video signals can be uninformative of each other. To limit the detrimental impact of faulty positives, we optimize a weighted contrastive learning loss, which down-weighs their contribution to the overall loss. Second, since self-supervised contrastive learning relies on random sampling of negative instances, instances that are semantically similar to the base instance can be used as faulty negatives. To alleviate the impact of faulty negatives, we propose to optimize an instance discrimination loss with a soft target distribution that estimates relationships between instances. We validate our contributions through extensive experiments on action recognition tasks and show that they address the problems of audio-visual instance discrimination and improve transfer learning performance.

We present a self-supervised learning approach to learn audio-visual representations from video and audio. Our method uses contrastive learning for cross-modal discrimination of video from audio and vice-versa. We show that optimizing for cross-modal discrimination, rather than within-modal discrimination, is important to learn good representations from video and audio. With this simple but powerful insight, our method achieves highly competitive performance when finetuned on action recognition tasks. Furthermore, while recent work in contrastive learning defines positive and negative samples as individual instances, we generalize this definition by exploring cross-modal agreement. We group together multiple instances as positives by measuring their similarity in both the video and audio feature spaces. Cross-modal agreement creates better positive and negative sets, which allows us to calibrate visual similarities by seeking within-modal discrimination of positive instances, and achieve significant gains on downstream tasks.

We introduce a novel self-supervised pretext task for learning representations from audio-visual content. Prior work on audio-visual representation learning leverages correspondences at the video level. Approaches based on audio-visual correspondence (AVC) predict whether audio and video clips originate from the same or different video instances. Audio-visual temporal synchronization (AVTS) further discriminates negative pairs originated from the same video instance but at different moments in time. While these approaches learn high-quality representations for downstream tasks such as action recognition, they completely disregard the spatial cues of audio and visual signals naturally occurring in the real world. To learn from these spatial cues, we tasked a network to perform contrastive audio-visual spatial alignment of 360\degree video and spatial audio. The ability to perform spatial alignment is enhanced by reasoning over the full spatial content of the 360\degree video using a transformer architecture to combine representations from multiple viewpoints. The advantages of the proposed pretext task are demonstrated on a variety of audio and visual downstream tasks, including audio-visual correspondence, spatial alignment, action recognition and video semantic segmentation. Dataset and code are available at https://github.com/pedro-morgado/AVSpatialAlignment.

The role of pose invariance in image recognition and retrieval is studied. A taxonomic classification of embeddings, according to their level of invariance, is introduced and used to clarify connections between existing embeddings, identify missing approaches, and propose invariant generalizations. This leads to a new family of pose invariant embeddings (PIEs), derived from existing approaches by a combination of two models, which follow from the interpretation of CNNs as estimators of class posterior probabilities: a view-to-object model and an object-to-class model. The new pose-invariant models are shown to have interesting properties, both theoretically and through experiments, where they outperform existing multiview approaches. Most notably, they achieve good performance for both 1) classification and retrieval, and 2) single and multiview inference. These are important properties for the design of real vision systems, where universal embeddings are preferable to task specific ones, and multiple images are usually not available at inference time. Finally, a new multiview dataset of real objects, imaged in the wild against complex backgrounds, is introduced. We believe that this is a much needed complement to the synthetic datasets in wide use and will contribute to the advancement of multiview recognition and retrieval.

Real-world applications of object recognition often require the solution of multiple tasks in a single platform. Under the standard paradigm of network fine-tuning, an entirely new CNN is learned per task, and the final network size is independent of task complexity. This is wasteful, since simple tasks require smaller networks than more complex tasks, and limits the number of tasks that can be solved simultaneously. To address these problems, we propose a transfer learning procedure, denoted NetTailor, in which layers of a pre-trained CNN are used as universal blocks that can be combined with small task-specific layers to generate new networks. Besides minimizing classification error, the new network is trained to mimic the internal activations of a strong unconstrained CNN, and minimize its complexity by the combination of 1) a soft-attention mechanism over blocks and 2) complexity regularization constraints. In this way, NetTailor can adapt the network architecture, not just its weights, to the target task. Experiments show that networks adapted to simple tasks, such as character or traffic sign recognition, become significantly smaller than those adapted to hard tasks, such as fine-grained recognition. More importantly, due to the modular nature of the procedure, this reduction in network complexity is achieved without compromise of either parameter sharing across tasks, or classification accuracy.

We introduce an approach to convert mono audio recorded by a 360° video camera into spatial audio, a representation of the distribution of sound over the full viewing sphere. Spatial audio is an important component of immersive 360° video viewing, but spatial audio microphones are still rare in current 360° video production. Our system consists of end-to-end trainable neural networks that separate individual sound sources and localize them on the viewing sphere, conditioned on multi-modal analysis from the audio and 360° video frames. We introduce several datasets, including one filmed ourselves, and one collected in-the-wild from YouTube, consisting of 360° videos uploaded with spatial audio. During training, ground truth spatial audio serves as self-supervision and a mixed down mono track forms the input to our network. Using our approach we show that it is possible to infer the spatial localization of sounds based only on a synchronized 360° video and the mono audio track.

The role of semantics in zero-shot learning is considered. The effectiveness of previous approaches is analyzed according to the form of supervision provided. While some learn semantics independently, others only supervise the semantic subspace explained by training classes. Thus, the former is able to constrain the whole space but lacks the ability to model semantic correlations. The latter addresses this issue but leaves part of the semantic space unsupervised. This complementarity is exploited in a new convolutional neural network (CNN) framework, which proposes the use of semantics as constraints for recognition. Although a CNN trained for classification has no transfer ability, this can be encouraged by learning an hidden semantic layer together with a semantic code for classification. Two forms of semantic constraints are then introduced. The first is a loss-based regularizer that introduces a generalization constraint on each semantic predictor. The second is a codeword regularizer that favors semantic-to-class mappings consistent with prior semantic knowledge while allowing these to be learned from data. Significant improvements over the state-of-the-art are achieved on several datasets.