We focus on better understanding the critical factors of augmentation-invariant representation learning. We revisit MoCo v2 and BYOL and try to prove the authenticity of the following assumption: different frameworks bring about representations of different characteristics even with the same pretext task. We establish the first benchmark for fair comparisons between MoCo v2 and BYOL, and observe: (i) sophisticated model configurations enable better adaptation to pre-training dataset; (ii) mismatched optimization strategies of pre-training and fine-tuning hinder model from achieving competitive transfer performances. Given the fair benchmark, we make further investigation and find asymmetry of network structure endows contrastive frameworks to work well under the linear evaluation protocol, while may hurt the transfer performances on long-tailed classification tasks. Moreover, negative samples do not make models more sensible to the choice of data augmentations, nor does the asymmetric network structure. We believe our findings provide useful information for future work.
论文检索
输入标题、作者或关键词，从 3,655 篇学术成果中精准定位
Video semantic segmentation has achieved great progress under the supervision of large amounts of labelled training data. However, domain adaptive video segmentation, which can mitigate data labelling constraints by adapting from a labelled source domain toward an unlabelled target domain, is largely neglected. We design temporal pseudo supervision (TPS), a simple and effective method that explores the idea of consistency training for learning effective representations from unlabelled target videos. Unlike traditional consistency training that builds consistency in spatial space, we explore consistency training in spatiotemporal space by enforcing model consistency across augmented video frames which helps learn from more diverse target data. Specifically, we design cross-frame pseudo labelling to provide pseudo supervision from previous video frames while learning from the augmented current video frames. The cross-frame pseudo labelling encourages the network to produce high-certainty predictions, which facilitates consistency training with cross-frame augmentation effectively. Extensive experiments over multiple public datasets show that TPS is simpler to implement, much more stable to train, and achieves superior video segmentation accuracy as compared with the state-of-the-art. Code is available at https://github.com/xing0047/TPS.
3D instance segmentation is a fundamental task for scene understanding, with a variety of applications in robotics and AR/VR. Many proposal-free methods have been proposed recently for this task, with remarkable results and high efficiency. However, these methods heavily rely on instance centroid regression and do not explicitly detect object boundaries, thus may mistakenly group nearby objects into the same clusters in some scenarios. In this paper, we define a novel concept of ”regional purity"" as the percentage of neighboring points belonging to the same instance within a fixed-radius 3D space. Intuitively, it indicates the likelihood of a point belonging to the boundary area. To evaluate the feasibility of predicting regional purity, we design a strategy to build a random scene toy dataset based on existing training data. Besides, using toy data is a ”free"" way of data augmentation on learning regional purity, which eliminates the burdens of additional real data. We propose Regional Purity Guided Network (RPGN), which has separate branches for predicting semantic class, regional purity, offset, and size. Predicted regional purity information is utilized to guide our clustering algorithm. Experimental results demonstrate that using regional purity can simultaneously prevent under-segmentation and over-segmentation problems during clustering.
Data augmentation is developed for increasing the amount and diversity of training data to enhance model learning. Compared to 2D images, point clouds, with the 3D geometric nature as well as the high collection and annotation costs, pose great challenges and potentials for augmentation. This paper presents a 3D augmentation method that explores the structural variance across multiple point clouds, and generates more diverse point clouds to enrich the training set. Specifically, we propose an attention module that decomposes a point cloud into several disjoint point subsets, called divisions, in a way where each division has a corresponding division in another point cloud. The augmented point clouds are synthesized by swapping matched divisions. They exhibit high diversity since both intra- and inter-cloud variations are explored, hence useful for downstream tasks. The proposed method for augmentation can act as a module and be integrated into a point-based network. The resultant framework is end-to-end trainable. The experiments show that it achieves state-of-the-art performance on the ModelNet40 and ModelNet10 benchmarks. The code for this work is publicly available.
This work considers supervised contrastive learning for semantic segmentation. We apply contrastive learning to enhance the discriminative power of the multi-scale features extracted by semantic segmentation networks. Our key methodological insight is to leverage samples from the feature spaces emanating from multiple stages of a model’s encoder itself requiring neither data augmentation nor online memory banks to obtain a diverse set of samples. To allow for such an extension we introduce an efficient and effective sampling process, that enables applying contrastive losses over the encoder’s features at multiple scales. Furthermore, by first mapping the encoder’s multi-scale representations to a common feature space, we instantiate a novel form of supervised local-global constraint by introducing cross-scale contrastive learning linking high-resolution local features to low-resolution global features. Combined, our multi-scale and cross-scale contrastive losses boost performance of various models (DeepLabv3, HRNet, OCRNet, UPerNet) with both CNN and Transformer backbones, when evaluated on 4 diverse datasets from natural (Cityscapes, PascalContext, ADE20K) but also surgical (CaDIS) domains. Our code is available at https://github.com/RViMLab/MS_CS_ContrSeg.
Collecting and labeling 3D data is costly. As a result, 3D resources for training are typically limited in quantity compared to the 2D images counterpart. In this work, we deal with the data scarcity challenge of 3D tasks by transferring knowledge from strong 2D models via abundant RGB-D images. Specifically, we utilize an strong and well-trained semantic segmentation model for 2D images to augment abundant RGB-D images with pseudo-label. The augmented dataset can then be used to pre-train 3D models. Finally, by simply fine-tuning on a few labeled 3D instances, our method already outperforms existing state-of-the-art that is tailored for 3D label efficiency. We further improve our results by using simple semi-supervised techniques ({\it i.e.}, mean teacher and entropy minimization). We verify the effectiveness of our pre-training on two popular 3D models on three different tasks. On ScanNet official evaluation, we establish new state-of-the-art semantic segmentation results on the data-efficient track.
Egocentric videos offer fine-grain information for high-fidelity modeling of human behaviors. Hands and interacting objects are one crucial aspect of understanding viewer’s behaviors and intentions. We provide a labeled dataset consisting of 11,235 egocentric images with per-pixel segmentation labels of hands and the interacting objects in diverse daily activities. Our dataset is the first to label detailed interacting hand-object contact boundaries. We introduce a context-aware compositional data augmentation technique to adapt to out-of-the-distribution YouTube egocentric video. We show that our robust hand-object segmentation model and dataset can serve as a foundation tool to boost or enable several downstream vision applications, such as: Hand state classification, video activity recognition, 3D mesh reconstruction of hand-object interaction, and Seeing through the hand with video inpainting in egocentric videos. All of our data and code will be released to the public.
Existing self-supervised monocular depth estimation methods can get rid of expensive annotations and achieve promising results. However, these methods suffer from severe performance degradation when directly adopting a model trained on a fixed resolution to evaluate at other different resolutions. In this paper, we propose a resolution adaptive self-supervised monocular depth estimation method (RA-Depth) by learning the scale invariance of the scene depth. Specifically, we propose a simple yet efficient data augmentation method to generate images with arbitrary scales for the same scene. Then, we develop a dual high-resolution network that uses the multi-path encoder and decoder with dense interactions to aggregate multi-scale features for accurate depth inference. Finally, to explicitly learn the scale invariance of the scene depth, we formulate a cross-scale depth consistency loss on depth predictions with different scales. Extensive experiments on the KITTI, Make3D and NYU-V2 datasets demonstrate that RA-Depth not only achieves state-of-the-art performance, but also exhibits a good ability of resolution adaptation.
Is It Necessary to Transfer Temporal Knowledge for Domain Adaptive Video Semantic Segmentation?
PDF ↗Video semantic segmentation is a fundamental and important task in computer vision, and it usually requires large-scale labeled data for training deep neural network models. To avoid laborious manual labeling, domain adaptive video segmentation approaches were recently introduced by transferring the knowledge from the source domain of self-labeled simulated videos to the target domain of unlabeled real-world videos. However, it leads to an interesting question -- while video-to-video adaptation is a natural idea, does the source data require to be videos? In this paper, we argue that it is not necessary to transfer temporal knowledge since the temporal continuity of video segmentation in the target domain can be estimated and enforced without reference to videos in the source domain. This motivates a new framework of Image-to-Video Domain Adaptive Semantic Segmentation (I2VDA), where the source domain is a set of images without temporal information. Under this setting, we bridge the domain gap via adversarial training based on only the spatial knowledge, and develop a novel temporal augmentation strategy, through which the temporal consistency in the target domain is well-exploited and learned. In addition, we introduce a new training scheme by leveraging a proxy network to produce pseudo-labels on-the-fly, which is very effective to improve the stability of adversarial training. Experimental results on two synthetic-to-real scenarios show that the proposed I2VDA method can achieve even better performance on video semantic segmentation than existing state-of-the-art video-to-video domain adaption approaches.
The promise of self-supervised learning (SSL) is to leverage large amounts of unlabeled data to solve complex tasks. While there has been excellent progress with simple, image-level learning, recent methods have shown the advantage of including knowledge of image structure. However, by introducing hand-crafted image segmentations to define regions of interest, or specialized augmentation strategies, these methods sacrifice the simplicity and generality that makes SSL so powerful. Instead, we propose a self-supervised learning paradigm that discovers this image structure by itself. Our method, Odin, couples object discovery and representation networks to discover meaningful image segmentations without any supervision. The resulting learning paradigm is simpler, less brittle, and more general, and achieves state-of-the-art transfer learning results for object detection and instance segmentation on COCO, and semantic segmentation on PASCAL and Cityscapes, while strongly surpassing supervised pre-training for video segmentation on DAVIS.
Fashion designs are rich in visual details associated with various visual attributes at both global and local levels. As a result, effective modeling and analyzing fashion requires fine-grained representations for individual attributes. In this work, we present a deep learning based online clustering method to jointly learn fine-grained fashion representations for all attributes at both instance and cluster level, where the attribute-specific cluster centers are online estimated. Based on the similarity between fine-grained representations and cluster centers, attribute-specific embedding spaces are further segmented into class-specific embedding spaces for fine-grained fashion retrieval. To better regulate the learning process, we design a three-stage learning scheme, to progressively incorporate different supervisions at both instance and cluster level, from both original and augmented data, and with ground-truth and pseudo labels. Experiments on FashionAI and DARN datasets in retrieval task demonstrated the efficacy of our method compared with competing baselines.
Defining positive and negative samples is critical for learning visual variations of the semantic classes in an unsupervised manner. Previous methods either construct positive sample pairs as different data augmentations on the same image (i.e., single-instance-positive) or estimate a class prototype by clustering (i.e., prototype-positive), both ignoring the relative nature of positive/negative concepts in the real world. Motivated by the ability of humans in recognizing relatively positive/negative samples, we propose the Relative Contrastive Loss (RCL) to learn feature representation from relatively positive/negative pairs, which not only learns more real world semantic variations than the single-instance-positive methods but also respects positive-negative relativeness compared with absolute prototype-positive methods. The proposed RCL improves the linear evaluation for MoCo v3 by \textbf{+2.0\%} on ImageNet. Code will be released publicly upon acceptance.
CutMix is a popular augmentation technique commonly used for training modern convolutional and transformer vision networks. It was originally designed to encourage Convolution Neural Networks (CNNs) to focus more on an image’s global context instead of local information, which greatly improves the performance of CNNs. However, we found it to have limited benefits for transformer-based architectures that naturally have a global receptive field. In this paper, we propose a novel data augmentation technique TokenMix to improve the performance of vision transformers. TokenMix mixes two images at token level via partitioning the mixing region into multiple separated parts. Besides, we show that the mixed learning target in CutMix, a linear combination of a pair of the ground truth labels, might be inaccurate and sometimes counter-intuitive. To obtain a more suitable target, we propose to assign the target score according to the content-based neural activation maps of the two images from a pre-trained teacher model, which does not need to have high performance. With plenty of experiments on various vision transformer architectures, we show that our proposed TokenMix helps vision transformers focus on the foreground area to infer the classes and enhances their robustness to occlusion, with consistent performance gains. Notably, we improve DeiT-T/S/B with +1% ImageNet top-1 accuracy. Besides, TokenMix enjoys longer training, which achieves 81.2% top-1 accuracy on ImageNet with DeiT-S trained for 400 epochs.
Deep neural networks have been the driving force behind the success in classification tasks, e.g., object and audio recognition. Impressive results and generalization have been achieved by a variety of recently proposed architectures, the majority of which are seemingly disconnected. In this work, we cast the study of deep classifiers under a unifying framework. In particular, we express state-of-the-art architectures (e.g., residual and non-local networks) in the form of different degree polynomials of the input. Our framework provides insights on the inductive biases of each model and enables natural extensions building upon their polynomial nature. The efficacy of the proposed models is evaluated on standard image and audio classification benchmarks. The expressivity of the proposed models is highlighted both in terms of increased model performance as well as model compression. Lastly, the extensions allowed by this taxonomy showcase benefits in the presence of limited data and long-tailed data distributions. We expect this taxonomy to provide links between existing domain-specific architectures.
Despite their impressive performance on image classification tasks, deep networks have a hard time generalizing to unforeseen corruptions of their data. To fix this vulnerability, prior works have built complex data augmentation strategies, combining multiple methods to enrich the training data. However, introducing intricate design choices or heuristics makes it hard to understand which elements of these methods are indeed crucial for improving robustness. In this work, we take a step back and follow a principled approach to achieve robustness to common corruptions. We propose PRIME, a general data augmentation scheme that relies on simple yet rich families of max-entropy image transformations. PRIME outperforms the prior art in terms of corruption robustness, while its simplicity and plug-and-play nature enable combination with other methods to further boost their robustness. We analyze PRIME to shed light on the importance of the mixing strategy on synthesizing corrupted images, and to reveal the robustness-accuracy trade-offs arising in the context of common corruptions. Finally, we show that the computational efficiency of our method allows it to be easily used in both on-line and off-line data augmentation schemes.
Class Subdivision (CCS) is important for many practical applications, where the training set originally annotated for a coarse class (e.g. bird) needs to further support its sub-classes recognition (e.g. swan, crow) with only very few fine-grained labeled samples. From the perspective of causal representation learning, these sub-classes inherit the same determinative factors of the coarse class, and their difference lies only in values. Therefore, to support the challenging CCS task with minimum fine-grained labeling cost, an ideal data augmentation method should generate abundant variants by manipulating these sub-class samples at the granularity of generating factors. For this goal, traditional data augmentation methods are far from sufficient. They often perform in highly-coupled image or feature space, thus can only simulate global geometric or photometric transformations. Leveraging the recent progress of factor-disentangled generators, Unbiased Manifold Augmentation (UMA) is proposed for CCS. With a controllable StyleGAN pre-trained for a coarse class, an approximate unbiased augmentation is conducted on the factor-disentangled manifolds for each sub-class, revealing the unbiased mutual information between the target sub-class and its determinative factors. Extensive experiments have shown that in the case of small data learning (less than 1\% fine-grained samples of commonly used), our UMA can achieve 10.37\% average improvement compared with existing data augmentation methods. On challenging tasks with severe bias, the accuracy is improved by up to 16.79\%.
The continual appearance of new objects in the visual world poses considerable challenges for current deep learning methods in real-world deployments. The challenge of new task learning is often exacerbated by the scarcity of data for the new categories due to rarity or cost. Here we explore the important task of Few-Shot Class-Incremental Learning (FSCIL) and its extreme data scarcity condition of one-shot. An ideal FSCIL model needs to perform well on all classes, regardless of their presentation order or paucity of data. It also needs to be robust to open-set real-world conditions and be easily adapted to the new tasks that always arise in the field. In this paper, we first reevaluate the current task setting and propose a more comprehensive and practical setting for the FSCIL task. Then, inspired by the similarity of the goals for FSCIL and modern face recognition systems, we propose our method --- Augmented Angular Loss Incremental Classification or ALICE. In ALICE, instead of the commonly used cross-entropy loss, we propose to use the angular penalty loss to obtain well-clustered features. As the obtained features not only need to be compactly clustered but also diverse enough to maintain generalization for future incremental classes, we further discuss how class augmentation, data augmentation, and data balancing affect classification performance. Experiments on benchmark datasets, including CIFAR100, miniImageNet, and CUB200, demonstrate the improved performance of ALICE over the state-of-the-art FSCIL methods.
Synthetic data is a scalable alternative to manual supervision, but it requires overcoming the sim-to-real domain gap. This discrepancy between virtual and real worlds is addressed by two seemingly opposed approaches: improving the realism of simulation or foregoing realism entirely via domain randomization. In this paper, we show that the recent progress in neural rendering enables a new unified approach we call Photo-realistic Neural Domain Randomization (PNDR). We propose to learn a composition of neural networks that acts as a physics-based ray tracer generating high-quality renderings from scene geometry alone. Our approach is modular, composed of different neural networks for materials, lighting, and rendering, thus enabling randomization of different key image generation components in a differentiable pipeline. Once trained, our method can be combined with other methods and used to generate photo-realistic image augmentations online and significantly more efficiently than via traditional ray-tracing. We demonstrate the usefulness of PNDR through two downstream tasks: 6D object detection and monocular depth estimation. Our experiments show that training with PNDR enables generalization to novel scenes and significantly outperforms the state of the art in terms of real-world transfer.
Machine learning classifiers are typically trained to minimise the average error across a dataset. Unfortunately, in practice, this process often exploits spurious correlations caused by subgroup imbalance within the training data, resulting in high average performance but highly variable performance across subgroups. Recent work to address this problem proposes model patching with CAMEL. This previous approach uses generative adversarial networks to perform intra-class inter-subgroup data augmentations, requiring (a) the training of a number of computationally expensive models and (b) sufficient quality of model’s synthetic outputs for the given domain. In this work, we propose RealPatch, a framework for simpler, faster, and more data-efficient data augmentation based on statistical matching. Our framework performs model patching by augmenting a dataset with real samples, mitigating the need to train generative models for the target task. We demonstrate the effectiveness of RealPatch on three benchmark datasets, CelebA, Waterbirds and a subset of iWildCam, showing improvements in worst-case subgroup performance and in subgroup performance gap in binary classification. Furthermore, we conduct experiments with the imSitu dataset with 211 classes, a setting where generative model-based patching such as CAMEL is impractical. We show that RealPatch can successfully eliminate dataset leakage while reducing model leakage and maintaining high utility. The code for RealPatch can be found at https://github.com/wearepal/RealPatch.
Active learning (AL) is a label-efficient technique for training deep models when only a limited labeled set is available and the manual annotation is expensive. Implicit semantic data augmentation (ISDA) effectively extends the limited amount of labeled samples and increases the diversity of labeled sets without introducing a noticeable extra computational cost. The scarcity of labeled instances and the huge annotation cost of unlabelled samples encourage us to ponder on the combination of AL and ISDA. A nature direction is a pipelined integration, which selects the unlabeled samples via acquisition function in AL for labeling and generates virtual samples by changing the selected samples to semantic transformation directions within ISDA. However, this pipelined combination would not guarantee the diversity of virtual samples. This paper proposes diversity-aware semantic transformation active learning, or DAST-AL framework, that looks ahead the effect of ISDA in the selection of unlabeled samples. Specifically, DAST-AL exploits expected partial model change maximization (EPMCM) to consider selected samples’ potential contribution of the diversity to the labeled set by leveraging the semantic transformation within ISDA when selecting the unlabeled samples. After that, DAST-AL can confidently and efficiently augment the labeled set by implicitly generating more diverse samples. The empirical results on both image classification and semantic segmentation tasks show that the proposed DAST-AL can slightly outperform the state-of-the-art AL approaches. Under the same condition, the proposed method takes less than 3 minutes for the first cycle of active labeling while the existing agreement discrepancy selection incurs more than 40 minutes.