Beyond Covariance: SICE and Kernel Based Visual Feature Representation

Zhang, Jianjia; Wang, Lei; Zhou, Luping; Li, Wanqing

doi:10.1007/s11263-020-01376-1

Cited by 22 publications

(12 citation statements)

References 89 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The Micchelli � s theorem [16] has proven that K i is guaranteed to be nonsingular no matter the values of feature dimensionality p and feature vector number hw. In summary, RBF kernel matrix K i could capture non-linear second order statistics and does not suffer from singular issue, the advantages of K i over covariance matrix C i have been corroborated theoretically and empirically in [66].…”

Section: Kernel Poolingmentioning

confidence: 74%

“…The covariance matrix is second order pooling in essence, which is able to learn the co-occurrence information between each pair of features. However, covariance matrix C i is type of linear kernel function [66], as a result, it is prone to be a singular matrix when feature dimension p is larger than feature vector number hw.…”

Section: Kernel Poolingmentioning

confidence: 99%

“…Accordingly, we follow [11,66] to employ non-linear kernel function RBF to capture the second order information between each pair of features. Let b k be the k-th row of B i ∈ ℝ p×hw , that is the k-th feature.…”

Section: Kernel Poolingmentioning

confidence: 99%

See 2 more Smart Citations

Kernel pooling feature representation of pre-trained convolutional neural networks for leaf recognition

Feng

2021

Multimed Tools Appl

View full text Add to dashboard Cite

Due to the presence of various types of factors, such as illumination, viewpoint, intra-class complexity, and inter-class similarity, which make plant leaf recognition still a challenging research problem. In this paper, we present a very simple, yet effective feature representation method for plant leaf recognition. Concretely, it comprises four stages. Firstly, each leaf image is fed into an imagenet pre-trained CNN model to extract activated feature maps in a specified layer. Secondly, inspired by 1 × 1 convolution, we exploit principle component analysis to learn the 1 × 1 convolution filters. As a result, it not only eliminates the redundant information, reduces the feature dimension adaptively that is beneficial to the subsequent high order pooling, but also increases classification accuracy. Thirdly, kernel pooling is employed to capture second order statistics between each pair of features with the purpose of learning more discriminative information. Finally, matrix sqrt and upper triangle are performed to obtain the final leaf representation, which is utilized for classification and retrieval by the euclidean distance based nearest neighbor classifier. Extensive experiments are conducted on four representative plant leaf datasets, Flavia, Swedish, MEW2012, ICL, to validate the effectiveness of our method. For classification task, our method achieves outstanding and better average classification accuracies than the comparative state-of-theart baselines. For retrieval task, our method gets significant higher or competitive MAP scores. Our implementation code will be available at https:// github. com/ fengs hu666 666/ leafr ecogn ition.

show abstract

Section: Kernel Poolingmentioning

confidence: 74%

Section: Kernel Poolingmentioning

confidence: 99%

See 1 more Smart Citation

Kernel pooling feature representation of pre-trained convolutional neural networks for leaf recognition

Feng

2021

Multimed Tools Appl

View full text Add to dashboard Cite

show abstract

“…( 9) is only in the covariance component. For the robust estimation methods of covariance, we adopt the iSQRT-COV rather than other powerful covariance estimation methods [29]- [32] due to its friendly performance on GPU and its original version considered in the Gaussian setting.…”

Section: Robust Distribution Knowledge Embeddingmentioning

confidence: 99%

Distribution Knowledge Embedding for Graph Pooling

Chen¹,

Song²,

Liu³

et al. 2021

Preprint

View full text Add to dashboard Cite

Graph-level representation learning is the pivotal step for downstream tasks that operate on the whole graph. The most common approach to this problem heretofore is graph pooling, where node features are typically averaged or summed to obtain the graph representations. However, pooling operations like averaging or summing inevitably cause massive information missing, which may severely downgrade the final performance. In this paper, we argue what is crucial to graphlevel downstream tasks includes not only the topological structure but also the distribution from which nodes are sampled. Therefore, powered by existing Graph Neural Networks (GNN), we propose a new plug-and-play pooling module, termed as Distribution Knowledge Embedding (DKEPool), where graphs are rephrased as distributions on top of GNNs and the pooling goal is to summarize the entire distribution information instead of retaining a certain feature vector by simple predefined pooling operations. A DKEPool network de facto disassembles representation learning into two stages, structure learning and distribution learning. Structure learning follows a recursive neighborhood aggregation scheme to update node features where structure information is obtained. Distribution learning, on the other hand, omits node interconnections and focuses more on the distribution depicted by all the nodes. Extensive experiments demonstrate that the proposed DKEPool significantly and consistently outperforms the state-of-the-art methods.

show abstract

“…Covariance descriptors have been extended to many other applications [83,58,87,17] including end-toend training of CNNs, leading to state-of-art results on action recognition, texture classification, scene and fine-grained recognition [34,35,42]. As second-order representations capture correlation patterns of features, they are a powerful tool used in several recognition pipelines [22,68,49,39,42,48,59,41,37,101].…”

Section: Introductionmentioning

confidence: 99%

Multi-Level Second-Order Few-Shot Learning

Zhang

Koniusz

2023

IEEE Trans. Multimedia

View full text Add to dashboard Cite

We propose a Multi-level Second-order (MlSo) fewshot learning network for supervised or unsupervised few-shot image classification and few-shot action recognition. We leverage so-called power-normalized second-order base learner streams combined with features that express multiple levels of visual abstraction, and we use self-supervised discriminating mechanisms. As Second-order Pooling (SoP) is popular in image recognition, we employ its basic element-wise variant in our pipeline. The goal of multi-level feature design is to extract feature representations at different layer-wise levels of CNN, realizing several levels of visual abstraction to achieve robust few-shot learning. As SoP can handle convolutional feature maps of varying spatial sizes, we also introduce image inputs at multiple spatial scales into MlSo. To exploit the discriminative information from multi-level and multiscale features, we develop a Feature Matching (FM) module that reweights their respective branches. We also introduce a selfsupervised step, which is a discriminator of the spatial level and the scale of abstraction. Our pipeline is trained in an end-to-end manner. With a simple architecture, we demonstrate respectable results on standard datasets such as Omniglot, mini-ImageNet, tiered-ImageNet, Open MIC, fine-grained datasets such as CUB Birds, Stanford Dogs and Cars, and action recognition datasets such as HMDB51, UCF101, and mini-MIT.

show abstract

Beyond Covariance: SICE and Kernel Based Visual Feature Representation

Cited by 22 publications

References 89 publications

Kernel pooling feature representation of pre-trained convolutional neural networks for leaf recognition

Kernel pooling feature representation of pre-trained convolutional neural networks for leaf recognition

Distribution Knowledge Embedding for Graph Pooling

Multi-Level Second-Order Few-Shot Learning

Contact Info

Product

Resources

About