According to the announcement, there are 2600 training images for task1 and task2. However, I found there are only 2595 training images in total. Besides, the corresponding attribute ground truths for task are 12971, which does not match with the number of training data (2595*5 = 12975). Is there something wrong?
The numbers expressed in our announcement email were rounded slightly, so I apologize for the confusion. I will update the Challenge website with more precise numbers, to clarify things for everyone.
Otherwise, the cardinality of training files is actually as follows:
Task 1:
2594 images
2594 ground truth masks
Task 2:
2594 images
12970 ground truth masks (2594 * 5 = 12970)
Task 3:
10015 images
1 ground truth CSV file. with 10015 data rows and 1 header row
Note that all of the ZIP bundles also include an ATTRIBUTION.txt and LICENSE.txt file, so the file counts will be greater by 2 accordingly.