Statistically-Motivated Second-Order Pooling

Yu, Kaicheng; Salzmann, Mathieu

doi:10.1007/978-3-030-01234-2_37

Cited by 40 publications

(23 citation statements)

References 41 publications

(108 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Four datasets are used: Describing Texture Dataset (DTD) (Cimpoi et al 2014), MINC-2500 (MINC) (Bell et al 2015), MIT-Indoor (Indoor) (Quattoni and Torralba 2009), and Caltech-UCSD Bird (CUB) (Xie et al 2013), which are the texture dataset, the material dataset, the indoor scene dataset, and the finegrained dataset, respectively. Following the work in (Yu and Salzmann 2018), the size of input images in DTD, Indoor, and CUB is 448 × 448, and the size of input images in MINC is 224 × 224. We use the VGG-16 network as the backbone, and layers after the conv5-3 layer are removed.…”

Section: Evaluation On the Image Classification Taskmentioning

confidence: 99%

Revisiting Bilinear Pooling: A Coding Perspective

Gao

Zhang

et al. 2020

AAAI

View full text Add to dashboard Cite

Bilinear pooling has achieved state-of-the-art performance on fusing features in various machine learning tasks, owning to its ability to capture complex associations between features. Despite the success, bilinear pooling suffers from redundancy and burstiness issues, mainly due to the rank-one property of the resulting representation. In this paper, we prove that bilinear pooling is indeed a similarity-based coding-pooling formulation. This establishment then enables us to devise a new feature fusion algorithm, the factorized bilinear coding (FBC) method, to overcome the drawbacks of the bilinear pooling. We show that FBC can generate compact and discriminative representations with substantially fewer parameters. Experiments on two challenging tasks, namely image classification and visual question answering, demonstrate that our method surpasses the bilinear pooling technique by a large margin.

show abstract

Section: Evaluation On the Image Classification Taskmentioning

confidence: 99%

Revisiting Bilinear Pooling: A Coding Perspective

Gao

Zhang

et al. 2020

AAAI

View full text Add to dashboard Cite

show abstract

“…Li et al [32] introduced a new structure to aggregate multiscale deep features to enhance feature representation ability and speed up experiment process for real-time semantic segmentation. To address these problem, Yu and Salzmann [33] proposed a parametric compression strategy to produce more compact representations than previous compression tactics. Gao et al [34] proposed NDDR layer to fuse single-task features by layerwise feature fusion for multitask feature learning.…”

Section: B Feature Fusionmentioning

confidence: 99%

Self-Layer and Cross-Layer Bilinear Aggregation for Fine-Grained Recognition in Cyber-Physical-Social Systems

et al. 2020

View full text Add to dashboard Cite

Cyber-Physical-Social Systems (CPSS) integrates cyber, physical and social spaces together, which makes our lives more convenient and intelligent by providing personalized service. In this paper, we will provide CPSS service for fine-grained recognition. Fine-grained visual recognition is a hot but challenging research in computer vision that aims to recognize object subcategories. The reason why it is challenging is that it extremely depends on the subtle discriminative features of local parts. Recently, some bilinear feature based methods were proposed, and the experimental results show state-of-the-art performance. However, most of them neglect the spatial relationships of part-region feature among multiple layers. In this paper, a novel approach of Self-layer and Cross-layer Bilinear Aggregation(SCBA) is proposed for fine-grained recognition. Firstly, a self-layer bilinear feature fusion module is proposed to model the spatial relationship of feature at the same layer. Secondly, we propose a cross-layer bilinear feature fusion module to capture the inter-layer interreaction of information to boost the ability of feature representation. In summary, the method we proposed not only can learn the correlations among different layers but the same layer, which makes it efficient and the experimental results show that it achieves state-of-the-art accuracy on three common fine-grained image datasets.

show abstract

“…However, using only first-order or second-order feature statistics as channel descriptor is limited in representing the global distribution of channel-wise feature responses, thus hindering the representational ability of CNNs. Meanwhile, recent works [3,4] have also shown that higher-order statistics are also helpful to improve discriminative ability of CNNs.…”

Section: Introductionmentioning

confidence: 99%

HOCA: Higher-Order Channel Attention for Single Image Super-Resolution

Dai

Chen

et al. 2021

ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

Convolutional neural networks (CNNs) have obtained great success in single image super-resolution (SR). More recent works (e.g., RCAN and SAN) have obtained remarkable performance with channel attention based on first-or second-order statistics of features. However, these methods neglect the rich feature statistics higher than second-order, thus hindering the representation ability of CNNs. To address this issue, we propose a higher-order channel attention (HOCA) module to enhance the representation ability of CNNs. In our HOCA module, to capture different types of semantic information, we first compute k-order of feature statistics, followed by channel attention to learn the feature interdependencies. Considering the diversity of input contents, we design a gate mechanism to adaptively select a specific k-order channel attention. Besides, our HOCA module serves as a plug-and-play module and can be easily plugged into existing state-of-art CNN-based SR methods. Extensive experiments on public benchmarks show that our HOCA module effectively improves the performance of various CNN-based SR methods.

show abstract

Statistically-Motivated Second-Order Pooling

Cited by 40 publications

References 41 publications

Revisiting Bilinear Pooling: A Coding Perspective

Revisiting Bilinear Pooling: A Coding Perspective

Self-Layer and Cross-Layer Bilinear Aggregation for Fine-Grained Recognition in Cyber-Physical-Social Systems

HOCA: Higher-Order Channel Attention for Single Image Super-Resolution

Contact Info

Product

Resources

About