Vision transformers (ViTs) have demonstrated remarkable performance in a variety of vision tasks. Despite their promising capabilities, training a ViT requires a large amount of diverse data. Several studies empirically found that using rich data augmentations, such as Mixup, Cutmix, and random erasing, is critical to the successful training of ViTs. Now, the use of rich data augmentations has become a standard practice in the current state. However, we report a vulnerability to this practice: Certain data augmentations such as Mixup cause a variance shift in the positional embedding of ViT, which has been a hidden factor that degrades the performance of ViT during the test phase. We claim that achieving a stable effect from positional embedding requires a specific condition on the image, which is often broken for the current data augmentation methods. We provide a detailed analysis of this problem as well as the correct configuration for these data augmentations to remove the side effects of variance shift. Experiments showed that adopting our guidelines improves the performance of ViTs compared with the current configuration of data augmentations.

L 2 regularization for weights in neural networks is widely used as a standard training trick. In addition to weights, the use of batch normalization involves an additional trainable parameter γ, which acts as a scaling factor. However, L 2 regularization for γ remains an undiscussed mystery and is applied in different ways depending on the library and practitioner. In this article, we study whether L 2 regularization for γ is valid. To explore this issue, we consider two approaches: (1) variance control to make the residual network behave like an identity mapping and (2) stable optimization through the improvement of effective learning rate. Through two analyses, we specify the desirable and undesirable γ to apply L 2 regularization and propose four guidelines for managing them. In several experiments, we observed that applying L 2 regularization to applicable γ increased 1% to 4% classification accuracy, whereas applying L 2 regularization to inapplicable γ decreased 1% to 3% classification accuracy, which is consistent with our four guidelines. Our proposed guidelines were further validated through various tasks and architectures, including variants of residual networks and transformers.

For the stable optimization of deep neural networks, regularization methods such as dropout and batch normalization have been used in various tasks. Nevertheless, the correct position to apply dropout has rarely been discussed, and different positions have been employed depending on the practitioners. In this study, we investigate the correct position to apply dropout. We demonstrate that for a residual network with batch normalization, applying dropout at certain positions increases the performance, whereas applying dropout at other positions decreases the performance. Based on theoretical analysis, we provide the following guideline for the correct position to apply dropout: apply one dropout after the last batch normalization but before the last weight layer in the residual branch. We provide detailed theoretical explanations to support this claim and demonstrate them through module tests. In addition, we investigate the correct position of dropout in the head that produces the final prediction. Although the current consensus is to apply dropout after global average pooling, we prove that applying dropout before global average pooling leads to a more stable output. The proposed guidelines are validated through experiments using different datasets and models.

Deep neural networks have been used in various fields, but their internal behavior in how they understand images is not well known. In this study, we discuss two counterintuitive properties of convolutional neural networks (CNNs). First, we evaluated the size of the receptive field of CNNs with their classification accuracy . Previous studies have attempted to increase the size of the receptive field for performance gain. However, we observed that some CNNs with a smaller receptive field can achieve higher classification accuracy . In this regard, we claim that a larger receptive field does not guarantee improved classification accuracy . Second, using the effective receptive field, we examined the contribution of each pixel to the output of CNN. Intuitively, each pixel is expected to equally contribute to the final output, but we found that there exist pixels in a partially dead state with little contribution to the output. We reveal that the reason for dead pixels lies in even stride operations with odd-sized kernels in CNN and propose a kernel padding method to remove the dead pixels. We demonstrated the vulnerability of CNNs with dead pixels when we detect a noise or small box that is on dead pixels. Our findings on dead pixels should be understood and considered in practical applications of CNN.

Transmission line (TL) inspection is important for ensuring a stable supply of electricity to rural areas. Currently, there are several TL detection approaches based on computer vision; however, they have limitations owing to background clutter in visible light images. This article presents a novel multimodal image feature fusion module that utilizes both visible light and infrared images to enhance the TL-detection performance. The proposed module consists of a multibranch feature extraction (MFE) block followed by a channelwise attention (CA) block. The first block extracts the representative features of each modal input using multiple branches. The outputs of the MFE block are jointly aggregated into an attention vector in the CA block. Finally, the attention vector recalibrates each input feature of the proposed module. To reduce the number of additional parameters due to the insertion of the module, we introduced a channel-shrink factor in the MFE block and utilized a$1\times {1}$convolution in the CA block. Comparison experiments with various augmented conditions of day, night, fog, and snow were conducted on a real-world dataset, which we constructed by visible light and infrared images. The results showed that the proposed module outperformed not only the case of single modal input but also the state-of-the-art fusion methods, regardless of the baseline networks. Additionally, the proposed module showed effectiveness in terms of capacity when the baseline network has a large number of weight parameters.