Thank you for your commendable efforts in your work. I have a question regarding the split of the Stanford Cars dataset, which comprises 16,185 images representing 196 car models, in comparison to other tasks.
In most metric-learning literature, the dataset split is described as follows: "The first 98 classes (8,054 images) are used for training, and the remaining 98 classes (8,131 images) are held out for testing."
However, the split mentioned in the Torchvision documentation states that "The data is split into 8,144 training images and 8,041 testing images, with an approximately 50-50 split for each class.", the training and testing split of which is different from your work.
Unfortunately, the official website is currently inaccessible, leaving me uncertain about the specific split used in this implementation.
Could you kindly provide me with a detailed split list (rather than the raw images) used in your implementation of the Stanford Cars dataset?
Thank you for your commendable efforts in your work. I have a question regarding the split of the Stanford Cars dataset, which comprises 16,185 images representing 196 car models, in comparison to other tasks.
In most metric-learning literature, the dataset split is described as follows: "The first 98 classes (8,054 images) are used for training, and the remaining 98 classes (8,131 images) are held out for testing."
However, the split mentioned in the Torchvision documentation states that "The data is split into 8,144 training images and 8,041 testing images, with an approximately 50-50 split for each class.", the training and testing split of which is different from your work.
Unfortunately, the official website is currently inaccessible, leaving me uncertain about the specific split used in this implementation.
Could you kindly provide me with a detailed split list (rather than the raw images) used in your implementation of the Stanford Cars dataset?
Thank you for your attention to this matter.