Flickr-Scenery

我们在Places365数据集和收集的Flickr风景数据集上进行实验。在补充材料中可以找到有关具有结构化样本 (例如建筑物) 的图像的更多结果。类似地，我们在Places365数据集的子集上评估我们的方法。我们从Places365数据集选择25个风景类，其中包含62,500个样本的子集。为了进一步分析我们方法的泛化，我们通过从Flickr收集54,710图像的大规模风景图像数据库来构建flickrscape数据集。所有图像均经过中心裁剪并调整大小为256 × 256像素。对于这两个数据集，我们将数据分成80% 、10% 进行训练、验证和测试。所有定量和定性实验仅在测试集中进行评估。源代码，经过训练的模型和Flickr-scines数据集将公开提供。We conducted experiments on the Places365 dataset and a Flickr landscape dataset we collected. Additional results for images with structured samples (e.g., buildings) can be found in the supplementary materials. Similarly, we evaluated our method on a subset of the Places365 dataset. We selected 25 landscape categories from the Places365 dataset, forming a subset containing 62,500 samples. To further analyze the generalization capability of our method, we constructed the flickrscape dataset, a large-scale landscape image database consisting of 54,710 images collected from Flickr. All images were center-cropped and resized to 256 × 256 pixels. For both datasets, we split the data into training, validation, and test sets at a ratio of 80%, 10%, and 10% respectively. All quantitative and qualitative experiments were evaluated exclusively on the test set. Source code, trained models, and the Flickr-scines dataset will be made publicly available.

搜集汇总

背景与挑战

背景概述

Flickr-Scenery是一个用于计算机视觉研究的公开数据集，基于Places365和Flickr收集的风景图像构建，包含25个风景类，图像均经过中心裁剪并调整为256×256像素。数据集分为训练、验证和测试集，旨在评估图像外绘方法的泛化能力，由多所大学和公司联合发布，相关资源已开源。