State-of-the-art object detection approaches typically rely on pre-trained classification models to achieve better performance and faster convergence. We hypothesize that classification pre-training strives to achieve translation invariance, and consequently ignores the localization aspect of the problem. We propose a new large-scale pre-training strategy for detection, where noisy class labels are available for all images, but not bounding-boxes. In this setting, we augment standard classification pre-training with a new detection-specific pretext task. Motivated by the noise-contrastive learning based self-supervised approaches, we design a task that forces bounding boxes with high-overlap to have similar representations in different views of an image, compared to non-overlapping boxes. We redesign Faster R-CNN modules to perform this task efficiently. Our experimental results show significant improvements over existing weakly-supervised and self-supervised pre-training approaches in both detection accuracy as well as fine-tuning speed.
论文检索
输入标题、作者或关键词，从 7,876 篇学术成果中精准定位
The nonlocal-based blocks are designed for capturing long-range spatial-temporal dependencies in computer vision tasks. Although having shown excellent performance, they still lack the mechanism to encode the rich, structured information among elements in an image or video. In this paper, to theoretically analyze the property of these nonlocal-based blocks, we provide a new perspective to interpret them, where we view them as a set of graph filters generated on a fully-connected graph. Specifically, when choosing the Chebyshev graph filter, a unified formulation can be derived for explaining and analyzing the existing nonlocal-based blocks (e.g., nonlocal block, nonlocal stage, double attention block). Furthermore, by concerning the property of spectral, we propose an efficient and robust spectral nonlocal block, which can be more robust and flexible to catch long-range dependencies when inserted into deep neural networks than the existing nonlocal blocks. Experimental results demonstrate the clear-cut improvements and practical applicabilities of our method on image classification, action recognition, semantic segmentation, and person re-identification tasks. Code are available at https://github.com/zh460045050/SNL_ICCV2021.
ECS-Net: Improving Weakly Supervised Semantic Segmentation by Using Connections Between Class Activation Maps
PDF ↗Image-level weakly supervised semantic segmentation is a challenging task. As classification networks tend to capture notable object features and are insensitive to overactivation, class activation map (CAM) is too sparse and rough to guide segmentation network training. Inspired by the fact that erasing distinguishing features force networks to collect new ones from non-discriminative object regions, we using relationships between CAMs to propose a novel weakly supervised method. In this work, we apply these features, learned from erased images, as segmentation supervision, driving network to study robust representation. In specifically, object regions obtained by CAM techniques are erased on images firstly. To provide other regions with segmentation supervision, Erased CAM Supervision Net (ECSNet) generates pixel-level labels by predicting segmentation results of those processed images. We also design the rule of suppressing noise to select reliable labels. Our experiments on PASCAL VOC 2012 dataset show that without data annotations except for ground truth image-level labels, our ECS-Net achieves 67.6% mIoU on test set and 66.6% mIoU on val set, outperforming previous state-of-the-art methods.
In few-shot learning scenarios, the challenge is to generalize and perform well on new unseen examples when only very few labeled examples are available for each task. Model-agnostic meta-learning (MAML) has gained the popularity as one of the representative few-shot learning methods for its flexibility and applicability to diverse problems. However, MAML and its variants often resort to a simple loss function without any auxiliary loss function or regularization terms that can help achieve better generalization. The problem lies in that each application and task may require different auxiliary loss function, especially when tasks are diverse and distinct. Instead of attempting to hand-design an auxiliary loss function for each application and task, we introduce a new meta-learning framework with a loss function that adapts to each task. Our proposed framework, named Meta-Learning with Task-Adaptive Loss Function (MeTAL), demonstrates the effectiveness and the flexibility across various domains, such as few-shot classification and few-shot regression.
We consider the problem of Federated Learning (FL) where numerous decentralized computational nodes collaborate with each other to train a centralized machine learning model without explicitly sharing their local data samples. Such decentralized training naturally leads to issues of imbalanced or differing data distributions among the local models and challenges in fusing them into a central model. Existing FL methods deal with these issues by either sharing local parameters or fusing models via online distillation. However, such a design leads to multiple rounds of inter-node communication resulting in substantial bandwidth consumption, while also increasing the risk of data leakage and consequent privacy issues. To address these problems, we propose a new distillation-based FL framework that can preserve privacy by design, while also consuming substantially less network communication resources when compared to the current state-of-the-art. Our framework engages in inter-node communication using only publicly available and approved datasets, thereby giving explicit privacy control to the user. To distill knowledge among the various local models, our framework involves a novel ensemble distillation algorithm that uses both final prediction as well as model attention. This algorithm explicitly considers the diversity among various local nodes while also seeking consensus among them. This results in a comprehensive technique to distill knowledge from various decentralized nodes. We demonstrate the various aspects and the associated benefits of our FL framework through extensive experiments that produce state-of-the-art results on both classification and segmentation tasks on natural and medical images.
Recent works have advanced the performance of self-supervised representation learning by a large margin. The core among these methods is intra-image invariance learning. Two different transformations of one image instance are considered as a positive sample pair, where various tasks are designed to learn invariant representations by comparing the pair. Analogically, for video data, representations of frames from the same video are trained to be closer than frames from other videos, i.e. intra-video invariance. However, cross-video relation has barely been explored for visual representation learning. Unlike intra-video invariance, ground-truth labels of cross-video relation is usually unavailable without human labors. In this paper, we propose a novel contrastive learning method which explores the cross-video relation by using cycle-consistency for general image representation learning. This allows to collect positive sample pairs across different video instances, which we hypothesize will lead to higher-level semantics. We validate our method by transferring our image representation to multiple downstream tasks including visual object tracking, image classification, and action recognition. We show significant improvement over state-of-the-art contrastive learning methods. Project page is available at https://happywu.github.io/cycle_contrast_video.
Few-shot Learning has been studied to mimic human visual capabilities and learn effective models without the need of exhaustive human annotation. Even though the idea of meta-learning for adaptation has dominated the few-shot learning methods, how to train a feature extractor is still a challenge. In this paper, we focus on the design of training strategy to obtain an elemental representation such that the prototype of each novel class can be estimated from a few labeled samples. We propose a two-stage training scheme, Partner-Assisted Learning (PAL), which first trains a partner encoder to model pair-wise similarities and extract features serving as soft-anchors, and then trains a main encoder by aligning its outputs with soft-anchors while attempting to maximize classification performance. Two alignment constraints from logit-level and feature-level are designed individually. For each few-shot task, we perform prototype classification. Our method consistently outperforms the state-of-the-art method on four benchmarks. Detailed ablation studies of PAL are provided to justify the selection of each component involved in training.
In real-life applications, machine learning models often face scenarios where there is a change in data distribution between training and test domains. When the aim is to make predictions on distributions different from those seen at training, we incur in a domain generalization problem. Methods to address this issue learn a model using data from multiple source domains, and then apply this model to the unseen target domain. Our hypothesis is that when training with multiple domains, conflicting gradients within each mini-batch contain information specific to the individual domains which is irrelevant to the others, including the test domain. If left untouched, such disagreement may degrade generalization performance. In this work, we characterize the conflicting gradients emerging in domain shift scenarios and devise novel gradient agreement strategies based on gradient surgery to alleviate their effect. We validate our approach in image classification tasks with three multi-domain datasets, showing the value of the proposed agreement strategy in enhancing the generalization capability of deep learning models in domain shift scenarios.
Modern recognition systems require large amounts of supervision to achieve accuracy. Adapting to new domains requires significant data from experts, which is onerous and can become too expensive. Zero-shot learning requires an annotated set of attributes for a novel category. Annotating the full set of attributes for a novel category proves to be a tedious and expensive task in deployment. This is especially the case when the recognition domain is an expert domain. We introduce a new field-guide-inspired approach to zero-shot annotation where the learner model interactively asks for the most useful attributes that define a class. We evaluate our method on classification benchmarks with attribute annotations like CUB, SUN, and AWA2 and show that our model achieves the performance of a model with full annotations at the cost of a significantly fewer number of annotations. Since the time of experts is precious, decreasing annotation cost can be very valuable for real-world deployment.
Ternary Neural Networks (TNNs) have received much attention due to being potentially orders of magnitude faster in inference, as well as more power efficient, than full-precision counterparts. However, 2 bits are required to encode the ternary representation with only 3 quantization levels leveraged. As a result, conventional TNNs have similar memory consumption and speed compared with the standard 2-bit models, but have worse representational capability. Moreover, there is still a significant gap in accuracy between TNNs and full-precision networks, hampering their deployment to real applications. To tackle these two challenges, in this work, we first show that, under some mild constraints, computational complexity of the ternary inner product can be reduced by 2x. Second, to mitigate the performance gap, we elaborately design an implementation-dependent ternary quantization algorithm. The proposed framework is termed Fast and Accurate Ternary Neural Networks (FATNN). Experiments on image classification demonstrate that our FATNN surpasses the state-of-the-arts by a significant margin in accuracy. More importantly, speedup evaluation compared with various precisions is analyzed on several platforms, which serves as a strong benchmark for further research.
One-stage long-tailed recognition methods improve the overall performance in a "seesaw" manner, i.e., either sacrifice the head's accuracy for better tail classification or elevate the head's accuracy even higher but ignore the tail. Existing algorithms bypass such trade-off by a multi-stage training process: pre-training on imbalanced set and fine-tuning on balanced set. Though achieving promising performance, not only are they sensitive to the generalizability of the pre-trained model, but also not easily integrated into other computer vision tasks like detection and segmentation, where pre-training of classifier solely is not applicable. In this paper, we propose a one-stage long-tailed recognition scheme, ally complementary experts (ACE), where the expert is the most knowledgeable specialist in a sub-set that dominates its training, and is complementary to other experts in the less-seen categories without disturbed by what it has never seen. We design a distribution-adaptive optimizer to adjust the learning pace of each expert to avoid over-fitting. Without special bells and whistles, the vanilla ACE outperforms the current one-stage SOTA method by 3 10% on CIFAR10-LT, CIFAR100-LT, ImageNet-LT and iNaturalist datasets. It is also shown to be the first one to break the "seesaw" trade-off by improving the accuracy of the majority and minority categories simultaneously in only one stage.
Convolutional Neural Networks (CNNs) use pooling to decrease the size of activation maps. This process is crucial to increase the receptive fields and to reduce computational requirements of subsequent convolutions. An important feature of the pooling operation is the minimization of information loss, with respect to the initial activation maps, without a significant impact on the computation and memory overhead. To meet these requirements, we propose SoftPool: a fast and efficient method for exponentially weighted activation downsampling. Through experiments across a range of architectures and pooling methods, we demonstrate that SoftPool can retain more information in the reduced activation maps. This refined downsampling leads to improvements in a CNN's classification accuracy. Experiments with pooling layer substitutions on ImageNet1K show an increase in accuracy over both original architectures and other pooling methods. We also test SoftPool on video datasets for action recognition. Again, through the direct replacement of pooling layers, we observe consistent performance improvements while computational loads and memory requirements remain limited.
In this paper, we look at the problem of few-shot image classification that aims to learn a classifier for previously unseen classes and domains from few labeled samples. Recent methods use various adaptation strategies for aligning their visual representations to new domains or select the relevant ones from multiple domain-specific feature extractors. In this work, we present URL, which learns a single set of universal visual representations by distilling knowledge of multiple domain-specific networks after co-aligning their features with the help of adapters and centered kernel alignment. We show that the universal representations can be further refined for previously unseen domains by an efficient adaptation step in a similar spirit to distance learning methods. We rigorously evaluate our model in the recent Meta-Dataset benchmark and demonstrate that it significantly outperforms the previous methods while being more efficient.
The choice of activation functions is crucial for modern deep neural networks. Popular hand-designed activation functions like Rectified Linear Unit(ReLU) and its variants show promising performance in various tasks and models. Swish, the automatically discovered activation function, outperforms ReLU on many challenging datasets. However, it has two main drawbacks. First, the tree-based search space is highly discrete and restricted, making it difficult to searching. Second, the sample-based searching method is inefficient, making it infeasible to find specialized activation functions for each dataset or neural architecture. To tackle these drawbacks, we propose a new activation function called Piecewise Linear Unit(PWLU), which incorporates a carefully designed formulation and learning method. It can learn specialized activation functions and achieves SOTA performance on large-scale datasets like ImageNet and COCO. For example, on ImageNet classification dataset, PWLU improves 0.9%/0.53%/1.0%/1.7%/1.0% top-1 accuracy over Swish for ResNet-18/ResNet-50/MobileNet-V2/MobileNet-V3/EfficientNet-B0. PWLU is also easy to implement and efficient at inference, which can be widely applied in real-world applications.
Multi-view projection methods have demonstrated their ability to reach state-of-the-art performance on 3D shape recognition. Those methods learn different ways to aggregate information from multiple views. However, the camera view-points for those views tend to be heuristically set and fixed for all shapes. To circumvent the lack of dynamism of current multi-view methods, we propose to learn those view-points. In particular, we introduce the Multi-View Transformation Network (MVTN) that regresses optimal view-points for 3D shape recognition, building upon advances in differentiable rendering. As a result, MVTN can be trained end-to-end along with any multi-view network for 3D shape classification. We integrate MVTN in a novel adaptive multi-view pipeline that can render either 3D meshes or point clouds. MVTN exhibits clear performance gains in the tasks of 3D shape classification and 3D shape retrieval without the need for extra training supervision. In these tasks, MVTN achieves state-of-the-art performance on ModelNet40, ShapeNet Core55, and the most recent and realistic ScanObjectNN dataset (up to 6 % improvement). Interestingly, we also show that MVTN can provide network robustness against rotation and occlusion in the 3D domain.
In prevalent knowledge distillation, logits in most image recognition models are computed by global average pooling, then used to learn to encode the high-level and task-relevant knowledge. In this work, we solve the limitation of this global logit transfer in this distillation context. We point out that it prevents the transfer of informative spatial information, which provides localized knowledge as well as rich relational information across contexts of an input scene. To exploit the rich spatial information, we propose a simple yet effective logit distillation approach. We add a local spatial pooling layer branch to the penultimate layer, thereby our method extends the standard logit distillation and enables learning of both finely-localized knowledge and holistic representation. Our proposed method shows favorable accuracy improvement against the state-of-the-art methods on several image classification datasets. We show that our distilled students trained on the image classification task can be successfully leveraged for object detection and semantic segmentation tasks; this result demonstrates our method's high transferability.
The recently-developed DETR approach applies the transformer encoder and decoder architecture to object detection and achieves promising performance. In this paper, we handle the critical issue, slow training convergence, and present a conditional cross-attention mechanism for fast DETR training. Our approach is motivated by that the cross-attention in DETR relies highly on the content embeddings for localizing the four extremities and predicting the box, which increases the need for high-quality content embeddings and thus the training difficulty. Our approach, named conditional DETR, learns a conditional spatial query from the decoder embedding for decoder multi-head cross-attention. The benefit is that through the conditional spatial query, each cross-attention head is able to attend to a band containing a distinct region, e.g., one object extremity or a region inside the object box. This narrows down the spatial range for localizing the distinct regions for object classification and box regression, thus relaxing the dependence on the content embeddings and easing the training. Empirical results show that conditional DETR converges 6.7x faster for the backbones R50 and R101 and 10x faster for stronger backbones DC5-R50 and DC5-R101. Code is available at https://github.com/Atten4Vis/ConditionalDETR.
We present Multiscale Vision Transformers (MViT) for video and image recognition, by connecting the seminal idea of multiscale feature hierarchies with transformer models. Multiscale Transformers have several channel-resolution scale stages. Starting from the input resolution and a small channel dimension, the stages hierarchically expand the channel capacity while reducing the spatial resolution. This creates a multiscale pyramid of features with early layers operating at high spatial resolution to model simple low-level visual information, and deeper layers at spatially coarse, but complex, high-dimensional features. We evaluate this fundamental architectural prior for modeling the dense nature of visual signals for a variety of video recognition tasks where it outperforms concurrent vision transformers that rely on large scale external pre-training and are 5-10 more costly in computation and parameters. We further remove the temporal dimension and apply our model for image classification where it outperforms prior work on vision transformers. Code is available at: https://github.com/facebookresearch/SlowFast.
Few-shot classification aims at classifying categories of a novel task by learning from just a few (typically, 1 to 5) labeled examples. An effective approach to few-shot classification involves a prior model trained on a large-sample base domain, which is then finetuned over the novel few-shot task to yield generalizable representations. However, task-specific finetuning is prone to overfitting due to the lack of enough training examples. To alleviate this issue, we propose a new finetuning approach based on contrastive learning that reuses unlabelled examples from the base domain in the form of distractors. Unlike the nature of unlabelled data used in prior works, distractors belong to classes that do not overlap with the novel categories. We demonstrate for the first time that the inclusion of such distractors can significantly boost few-shot generalization. Our technical novelty includes a stochastic pairing of examples sharing the same category in the few-shot task and a weighting term that controls the relative influence of task-specific negatives and distractors. An important aspect of our finetuning objective is that it is agnostic to distractor labels and hence applicable to various base domain settings. More precisely, compared to state-of-the-art approaches, our method shows accuracy gains of up to 12% in cross-domain and up to 5% in unsupervised prior-learning settings. Our code is available at https://github.com/quantacode/Contrastive-Finetuning.git
Transformers have been recently adapted for large scale image classification, achieving high scores shaking up the long supremacy of convolutional neural networks. However the optimization of vision transformers has been little studied so far. In this work, we build and optimize deeper transformer networks for image classification. In particular, we investigate the interplay of architecture and optimization of such dedicated transformers. We make two architecture changes that significantly improve the accuracy of deep transformers. This leads us to produce models whose performance does not saturate early with more depth, for instance we obtain 86.5% top-1 accuracy on Imagenet when training with no external data, we thus attain the current sate of the art with less floating-point operations and parameters. Our best model establishes the new state of the art on Imagenet with Reassessed labels and Imagenet-V2 / match frequency, in the setting with no additional training data. We share our code and models