The iNaturalist Species Classification and Detection Dataset-Supplementary Material
The iNaturalist Species Classification and Detection Dataset-Supplementary Material
复制标题
iNaturalist 物种分类和检测数据集 - 补充材料
DOI：
--
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
Hartwig Adam
中科院分区：
文献类型：
--
作者：
Grant Van Horn;Oisin Mac Aodha;Yang Song;Yin Cui;Chenghe Sun;Alexander Shepard;Hartwig Adam
我们进行了一项实验，以了解现实世界中动物的大小和预测准确性之间是否存在任何关系。使用鸟类[4]和哺乳动物[2]身体大小的现有记录，我们在iNat2017中为与这些数据集重叠的每个类别分配了质量。对于特定物种，数量会因特定个体的生命阶段或性别而有所不同。在这里，我们只取平均值。这导致了795个物种的数据，从小的艾伦蜂鸟(小蜂鸟)到大型座头鲸Megaptera novaeangliae。在图1中，我们可以看到，中位精度随着物种质量的增加而降低。这些结果是初步的，但加强了这样的观察，即人类要为较大的哺乳动物拍摄良好的照片可能是具有挑战性的。对这些失败案例的更多分析可能会让我们在iNaturist上为摄影师提供更好的、针对特定物种的说明。
We performed an experiment to understand if there was any relationship between real world animal size and prediction accuracy. Using existing records for bird [4] and mammal [2] body sizes we assigned a mass to each of the classes in iNat2017 that overlapped with these datasets. For a given species, mass will vary due to the life stage or gender of the particular individual. Here, we simply take the average value. This resulted in data for 795 species, from the small Allen’s hummingbird (Selasphorus sasin) to the large Humpback whale Megaptera novaeangliae. In Fig. 1 we can see that median accuracy decreases as the mass of the species increases. These results are preliminary, but reinforce the observation that it can be challenging for humans to take good photographs of larger mammals. More analysis of these failure cases may allow us to produce better, species-specific, instructions for the photographers on iNaturalist.