Hi @YonghaoHe
The results between the paper and GitHub model zoo differ from:
We use FixedAP to compute the AP with a maximum of 10k predictions and report the results in our paper. In the GitHub repo, we use the naive AP as the metric.
The checkpoints in GitHub are trained based on the improved version and YOLO-World-v2 performs well than the paper version.
I found the precision gap of yolo-world-v1 between git and paper. See below: Do I miss something?