Attention-Based Dropout Layer for Weakly Supervised Single Object Localization and Semantic Segmentation

Choe, Jong‐In; Lee, Seungho; Shim, Hyunjung

doi:10.1109/tpami.2020.2999099

Cited by 83 publications

(55 citation statements)

References 52 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Furthermore, we have combined PSE block and SE block in the model for better performance. In the future, we will study how to use PSE blocks for more complex tasks [46][47][48], and use PSigmoid in other networks [49][50][51][52].…”

Section: Resultsmentioning

confidence: 99%

PSigmoid: Improving squeeze-and-excitation block with parametric sigmoid

et al. 2021

View full text Add to dashboard Cite

Squeeze-and-Excitation (SE) Networks won the last ImageNet Large-Scale Visual Recognition Challenge (ILSVRC) classification competition and is very popular in today's vision community. The SE block is the core of Squeeze-and-Excitation Network (SENet), which adaptively recalibrates channel-wise features and suppresses less useful ones. Since SE blocks can be directly used in existing models and effectively improve performance, SE blocks are widely used in a variety of tasks. In this paper, we propose a novel Parametric Sigmoid (PSigmoid) to enhance the SE block. We named the new module PSigmoid SE (PSE) block. The PSE block can not only suppress features in a channel-wise manner, but also enhance features. We evaluate the performance of our method on four common datasets including CIFAR-10, CIFAR-100, SVHN and Tiny ImageNet. Experimental results show the effectiveness of our method. In addition, we compare the differences between the PSE block and the SE block through a detailed analysis of the configuration. Finally, we use a combination of PSE block and SE block to obtain better performance.

show abstract

Section: Resultsmentioning

confidence: 99%

PSigmoid: Improving squeeze-and-excitation block with parametric sigmoid

et al. 2021

View full text Add to dashboard Cite

show abstract

“…Recently, various studies [36][37][38][39][40] have used the self-attention mechanism to improve the classification accuracy. Hu et al [36] proposed the squeeze-and-excitation (SE) block that increased the accuracy of the classification model based on the use of a one-dimensional (1D) channel self-attention map.…”

Section: Attention Mechanismmentioning

confidence: 99%

“…Wang et al [37] formulated self-attention as a non-local operation, covering the entire image region in one operation to model spatial-temporal dependencies in video sequences. Park et al [38] and Choe et al [40] proposed the bottleneck attention module (BAM) and attention-based dropout layer (ADL) that respectively produced spatial self-attention and importance maps with auxiliary convolutional layers. The produced self-attention map is applied to the input feature map to emphasize the object region.…”

Section: Attention Mechanismmentioning

confidence: 99%

See 1 more Smart Citation

SSD-EMB: An Improved SSD Using Enhanced Feature Map Block for Object Detection

Choi

Lee

Kang

et al. 2021

Sensors

View full text Add to dashboard Cite

The development of deep learning has achieved great success in object detection, but small object detection is still a difficult and challenging task in computer vision. To address the problem, we propose an improved single-shot multibox detector (SSD) using enhanced feature map blocks (SSD-EMB). The enhanced feature map block (EMB) consists of attention stream and feature map concatenation stream. The attention stream allows the proposed model to focus on the object regions rather than background owing to channel averaging and the effectiveness of the normalization. The feature map concatenation stream provides additional semantic information to the model without degrading the detection speed. By combining the output of these two streams, the enhanced feature map, which improves the detection of a small object, is generated. Experimental results show that the proposed model has high accuracy in small object detection. The proposed model not only achieves good detection accuracy, but also has a good detection speed. The SSD-EMB achieved a mean average precision (mAP) of 80.4% on the PASCAL VOC 2007 dataset at 30 frames per second on an RTX 2080Ti graphics processing unit, an mAP of 79.9% on the VOC 2012 dataset, and an mAP of 26.6% on the MS COCO dataset.

show abstract

“…Since both ReLU and Dropout have the function of zeroing or discarding a certain feature, we call their function here as selective feature transformation. "Attention" mechanism which is a popular technique uses weight to amplify or weaken the value of certain input variables or the block of feature map [27], and the feature selection method we proposed LSTM-RDN which has ReLU and Dropout layers also has similar select functions as Attention mechanism.…”

Section: Lstm-based Selective Feature Transformation Networkmentioning

confidence: 99%

One-dimensional DCNN Feature Selective Transformation with LSTM-RDN for Image Classification

Li¹,

Huang²,

Huang³

et al. 2021

Preprint

View full text Add to dashboard Cite

Feature selection and transformation are the important techniques in machine learning field. A good feature selection / transformation method will greatly improve the performance of classification algorithm. In this work, we proposed a simple but efficient image classification method which is based on two-stage processing strategy. In the first processing stage, the one-dimensional features are obtained from image by transfer learning with the pre-trained Deep Convolutional Neural Networks (DCNN). These one-dimensional DCNN features still have the shortcomings of information redundancy and weak distinguishing ability. Therefore, it is necessary to use feature transformation to continue to obtain more distinguishable features. We propose a feature learning and selective transformation network based on Long Short-Term Memory (LSTM) combing ReLU and Dropout layers (called LSTM-RDN) to further process DCNN one-dimensional features. . The verification experiments were conducted on three public object image datasets(Cifar10, Cifar100 and Fashion-MNIST), three fine-grained image datasets(CUB200-2011, Stanford-Cars, FGVC-Aircraft) and a COVID-19 dataset. In the experiments, we used several backbone network models, including AlexNet, VGG16, ResNet18, ResNet101, InceptionV2 and EfficientNet-b0. Experimental results have shown that through feature selective transformation, the recognition accuracy of these DCNN models can significantly exceed the classification accuracies of the state-of-the-art methods.

show abstract

Attention-Based Dropout Layer for Weakly Supervised Single Object Localization and Semantic Segmentation

Cited by 83 publications

References 52 publications

PSigmoid: Improving squeeze-and-excitation block with parametric sigmoid

PSigmoid: Improving squeeze-and-excitation block with parametric sigmoid

SSD-EMB: An Improved SSD Using Enhanced Feature Map Block for Object Detection

One-dimensional DCNN Feature Selective Transformation with LSTM-RDN for Image Classification

Contact Info

Product

Resources

About