iNaturalist Fine-Grained Geolocation
收藏资源简介：
细粒度识别区分具有细微视觉差异的类别。为了区分这些具有挑战性的视觉类别，利用其他信息是有帮助的。地理定位是丰富的附加信息来源，可用于提高细粒度分类准确性，但尚未得到充分研究。我们对这一领域的贡献是双重的。首先，据我们所知，这是第一篇论文，该论文系统地研究了通过使用地理位置先验，后处理或特征调制将地理位置信息整合到细粒度图像分类中的各种方法。其次，为了克服没有细粒度数据集具有完整地理位置信息的情况，我们通过向现有的流行数据集iNaturalist和YFCC100M提供补充信息，发布了两个具有地理位置的细粒度数据集。通过利用地理位置信息，我们将iNaturalist的top-1准确性从70.1% 提高到79.0%，以实现强大的仅基线图像模型。比较几种模型，我们发现最佳性能是通过后处理模型实现的，该模型消耗了仅图像基线的输出以及地理位置。但是，对于资源受限的模型 (MobileNetV2)，使用在像素和地理位置上联合训练的特征调制模型，性能更好: 精度从59.6% 提高到72.2%。我们的工作为将地理位置信息纳入服务器和设备上的细粒度识别模型提供了有力的依据。
Fine-grained recognition aims to distinguish categories with subtle visual differences. To differentiate these challenging visual categories, leveraging additional auxiliary information proves beneficial. Geolocation serves as a rich source of supplementary information that can enhance the accuracy of fine-grained classification, yet it remains understudied. Our contributions to this field are two-fold. First, to the best of our knowledge, this is the first paper that systematically investigates various approaches for integrating geolocation information into fine-grained image classification via geolocation priors, post-processing, or feature modulation. Second, to address the absence of complete geolocation information in existing fine-grained datasets, we release two geolocation-augmented fine-grained datasets by supplementing popular existing datasets iNaturalist and YFCC100M with additional geolocation metadata. By leveraging geolocation information, we improve the top-1 accuracy of the iNaturalist dataset from 70.1% to 79.0% on a strong image-only baseline model. Through comparative experiments across multiple models, we find that the best performance is achieved by the post-processing model, which takes both the outputs of the image-only baseline and geolocation information as inputs. However, for resource-constrained models such as MobileNetV2, the feature modulation model jointly trained on pixel inputs and geolocation information delivers better performance, boosting accuracy from 59.6% to 72.2%. Our work provides strong evidence for incorporating geolocation information into fine-grained recognition models deployed on both servers and edge devices.
