tjddus9597
A Question about your Nearest Neighbour Minibatch!
Thanks for Your Sharing! I want to follow your work to do further research! And I have a question about your NN batch contruction. Assume I use the Stanford Car dataset, it own 8144 images in training, and after I use your method, I found I get 8054 images in dataloader, can you explain the reason or just give me more details about that, I'm really intreseted in it!
Hi, thanks for your interest in our work!
The training and test splits indicated in the Cars dataset differ from those employed in metric learning. The dataset's annotated split pertains to classification, whereas in metric learning, it is partitioned based on class indices. Specifically, within Cars-196 dataset, metric learning utilizes the first 98 classes for training, encompassing a total of 8,054 images. Thus, our code appears to have loaded data accurately.
If you would like to know more about splitting the dataset, please see the Lifted Structure paper (https://arxiv.org/pdf/1511.06452.pdf)!
