Class-Balanced Loss Based on Effective Number of Samples

Cui, Yin; Jia, Menglin; Lin, Tsung-Yi; Song, Yang; Belongie, Serge

doi:10.48550/arxiv.1901.05555

Cited by 42 publications

(53 citation statements)

References 32 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This situation is known in data science as a problem affected by "class unbalanced" data. To deal with this we use the above loss function which has been shown to be well suited for segmentation problems that are affected by class unbalanced data (Cui et al 2019). We further used the Adaptive Moment Estimator Adam (Kingma & Ba 2014), an optimized stochastic gradient descent algorithm for error minimization.…”

Section: Our Network Segu-netmentioning

confidence: 99%

Deep learning approach for identification of HII regions during reionization in 21-cm observations

Bianco,

Giri,

Iliev

et al. 2021

Preprint

View full text Add to dashboard Cite

The upcoming Square Kilometre Array (SKA-Low) will map the distribution of neutral hydrogen during reionization, and produce a tremendous amount of 3D tomographic data. These images cubes will be subject to instrumental limitations, such as noise and limited resolution. Here we present SegU-Net, a stable and reliable method for identification of neutral and ionized regions in these images. SegU-Net is a U-Net architecture based convolutional neural network (CNN) for image segmentation. It is capable of segmenting our image data into meaningful features (ionized and neutral regions) with greater accuracy compared to previous methods. We can estimate the true ionization history from our mock observation of SKA with an observation time of 1000 h with more than 87 per cent accuracy. We also show that SegU-Net can be used to recover various topological summary statistics, such as size distributions and Betti numbers, with a relative difference of only a few per cent. These summary statistics characterise the non-Gaussian nature of the reionization process.

show abstract

Section: Our Network Segu-netmentioning

confidence: 99%

Deep learning approach for identification of HII regions during reionization in 21-cm observations

Bianco,

Giri,

Iliev

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…A pruning hyper-parameter p is established and set to 4 which allows the network to decide the number of filters to remove before the first stage of depthwise separable convolution. A class balanced focal loss function [7] is also added to assist and ease the effect of this pruning process.…”

Section: A Widely Popular Model Compression Technique Calledmentioning

confidence: 99%

Image Classification with CondenseNeXt for ARM-Based Computing Platforms

Kalgaonkar

El-Sharkawy

2021

2021 IEEE International IOT, Electronics and Mechatronics Conference (IEMTRONICS)

View full text Add to dashboard Cite

In this paper, we demonstrate the implementation of our ultra-efficient deep convolutional neural network architecture: CondenseNeXt on NXP BlueBox, an autonomous driving development platform developed for self-driving vehicles. We show that CondenseNeXt is remarkably efficient in terms of FLOPs, designed for ARM-based embedded computing platforms with limited computational resources and can perform image classification without the need of a CUDA enabled GPU. CondenseNeXt utilizes the state-of-the-art depthwise separable convolution and model compression techniques to achieve a remarkable computational efficiency.Extensive analyses are conducted on CIFAR-10, CIFAR-100 and ImageNet datasets to verify the performance of Con-denseNeXt Convolutional Neural Network (CNN) architecture. It achieves state-of-the-art image classification performance on three benchmark datasets including CIFAR-10 (4.79% top-1 error), CIFAR-100 (21.98% top-1 error) and ImageNet (7.91% single model, single crop top-5 error). CondenseNeXt achieves final trained model size improvement of 2.9+ MB and up to 59.98% reduction in forward FLOPs compared to CondenseNet and can perform image classification on ARM-Based computing platforms without needing a CUDA enabled GPU support, with outstanding efficiency.

show abstract

“…The way of our PSIS method to enhance instance balance is related to works those handle the issue of class imbalance in image classification [4,8,11,22,26,48]. Among them, methods [3,4,22] balance distribution of classes by re-sampling for those containing less training samples.…”

Section: Related Workmentioning

confidence: 99%

“…We can clearly see that distribution of instances is longtailed and detection performance of each class is quite different. Many works demonstrate data imbalance brings side effect on image classification [7,8,11,26,29,48]. Meanwhile, detection difficulties of different classes greatly vary, and distribution of instances and detection performance of each class do not always behave in the same way.…”

Section: Introductionmentioning

confidence: 99%

“…To this end, we explore a progressive and selective scheme for our IS strategy to generate synthetic images for object detection while training detectors with a class-balanced loss. The resulted method is called Progressive and Selective Instance-Switching (PSIS), which enhances instance balance by selectively performing IS (i.e., inversely proportional to the original instance frequency) for adjusting distribution of instances [3,4,22] and combining with a class-balanced loss [8]. To consider class importance, PSIS augments training dataset in a progressive manner, which is guided by detection performance (i.e., increase of times of IS for classes with lowest APs).…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Data Augmentation for Object Detection via Progressive and Selective Instance-Switching

Wang,

Yang

et al. 2019

Preprint

View full text Add to dashboard Cite

Collection of massive well-annotated samples is effective in improving object detection performance but is extremely laborious and costly. Instead of data collection and annotation, the recently proposed Cut-Paste methods [12,15] show the potential to augment training dataset by cutting foreground objects and pasting them on proper new backgrounds. However, existing Cut-Paste methods cannot guarantee synthetic images always precisely model visual context, and all of them require external datasets. To handle above issues, this paper proposes a simple yet effective instance-switching (IS) strategy, which generates new training data by switching instances of same class from different images. Our IS naturally preserves contextual coherence in the original images while requiring no external dataset. For guiding our IS to obtain better object performance, we explore issues of instance imbalance and class importance in datasets, which frequently occur and bring adverse effect on detection performance. To this end, we propose a novel Progressive and Selective Instance-Switching (PSIS) method to augment training data for object detection. The proposed PSIS enhances instance balance by combining selective re-sampling with a class-balanced loss, and considers class importance by progressively augmenting training dataset guided by detection performance. The experiments are conducted on the challenging MS COCO benchmark, and results demonstrate our PSIS brings clear improvement over various state-of-the-art detectors (e.g., Faster R-CNN, FPN, Mask R-CNN and SNIPER), showing the superiority and generality of our PSIS. Code and models are available at: https://github.com/Hwang64/PSIS.

show abstract

Class-Balanced Loss Based on Effective Number of Samples

Cited by 42 publications

References 32 publications

Deep learning approach for identification of HII regions during reionization in 21-cm observations

Deep learning approach for identification of HII regions during reionization in 21-cm observations

Image Classification with CondenseNeXt for ARM-Based Computing Platforms

Data Augmentation for Object Detection via Progressive and Selective Instance-Switching

Contact Info

Product

Resources

About