Open zhangzx-123 opened 4 months ago
The yolo score is not that easy reproduced like FID, as we need to first train a detection model on some kind of dataset with specific configs and then test. However, the training setting and config is complicated. Therefore, I recommend the followers to apply some open vocabulary public trained detection model (such as grounding dino) for easy reproduction. As we have shared the pretrained weights of our LayoutDiffusion, you can easily test again on our model with your own evaluation setting.
I want to evaluate the yolo score of the images generated using coco-staff data with 171 kinds of object. The code of LAMA can only detect 80 kinds of object. Could you give me the evaluate code and yolo weights? thanks a lot!!!