- Bounding box annotations: 2.07M (1.24M in train, 830K in validation+test)
- Human-verified image-level labels in training set: 9.38M (4.30M positive examples, 5.07M negative examples)
- Cleanup of human-verifications in the validation+test sets
- The V1 validation set is now partitioned into validation and test sets.