Occlusions for Effective Data Augmentation in Image Classification

Abstract

Deep networks for visual recognition are known to leverage "easy to recognise" portions of objects such as faces and distinctive texture patterns. The lack of a holistic understanding of objects may increase fragility and overfitting. In recent years, several papers have proposed to address this issue by means of occlusions as a form of data augmentation. However, successes have been limited to tasks such as weak localization and model interpretation, but no benefit was demonstrated on image classification on large-scale datasets. In this paper, we show that, by using a simple technique based on batch augmentation, occlusions as data augmentation can result in better performance on ImageNet for high-capacity models (e.g., ResNet50). We also show that varying amounts of occlusions used during training can be used to study the robustness of different neural network architectures.

One-Stream Vision-Language Memory Network for Object Tracking

“…Task Settings Zero-shot Few-shot TF Few-shot TPT [51] and [38,40,43,46,75] DiffTPT [12] CoOp [74] and [3,22,36,50,59,63,65,66,73,76] Tip-Adapter [68], and [77] SuS-X [56] CALIP [14] CaFo [70] DMN (Ours) Table 1. Summary of adaptation methods for vision-language models.…”

“…1. In the zero-shot setting where labeled training data are unavailable, one primary research direction is how to extract richer information from the test samples [12,14,51,75] and class names [12,38,40,43,56].…”

“…For the former group, CALIP [14] enhances feature extraction through attention mechanisms, and instance-adaptive prompts are explored using consistency regularization in [12,51]. Leveraging class names, some approaches [56,70] generate synthetic training samples utilizing additional image generation models [7,47], and others [38,43,46] craft advanced text prompts by querying pre-trained large language models.…”

Prior distributions-based data augmentation for object detection

“…The traditional data augmentation methods used to adopt geometric transformation of the image, such as horizontal flip (Ren et al, 2015), multi-scale strategy (Singh et al, 2018), patch cropping (Liu et al, 2016), and colour change of the image (Perez and Wang, 2017). Recently, more complicated image transformations are proposed, such as randomly erasing image (Zhong et al, 2020), blocking a part of image, and mixing images (Fong, 2019;Yun et al, 2019). The above methods mostly focus on the global transformation of image, while the object detection task pays more attention to the local features of the foreground instances in the image.…”

“…To tackle the problem, various of data augmentation methods are widely used. For instance, many methods (Zhong et al, 2020;Fong, 2019;Yun et al, 2019) have been proposed in image classification task while they have limitations in the object detection task. Different from the image classification task, in addition to classifing instances in the image, object detection task needs to label the locations of them.…”