Mazin-Hnewa
commented
2 years ago In the paper for cityscapes->foggy citycapes, we reported the result based on 5 popular classes and the size of the input test images is 416 X 416.
In the code, the result is based on 8 classes and the size of the input test images is 608 X 608. This is why the results are different between paper and code.
For BDD. we manually selected the rainy images from the dataset because many images are tagged as rainy but they are cloudy or sunny.
This means there is an overlap between sunny and rainy images in the original BDD. You can check the dataset by yourself. This is the reason that you found that the original yolov4 outperforms ms-dayolo with BDD.
I will provide the selected filtered data later.
Also, which other domain shift scenarios you tested that the original yolov4 outperforms ms-dayolo? Can you share the resutls?
For model weights in the paper, some files are lost because my hard disk is dead. So I retrain the model and get slightly different results because every time you train YOLOv4, you will get slightly different results as YOLOv4 selects random images in each iteration durning training. cityscapes->foggy citycapes: yolov4, MS-DAYOLO. BDD sunny -> rainy: yolov4, MS-DAYOLO
Thanks to the author for the code sharing.
The author's code was used to train original yolov4 achieves 38.0% mAP for 8 classes in cityscapes->foggy citycapes, which is significantly different from the data on the paper. Also, ms-dayolo was used to train on BDD from sunny to rainy, it is found that the original yolov4 outperform ms-dayolo. In fact, I tested code provided by the author and found that the original yolov4 outperforms ms-dayolo in many domain shift scenarios.
Are there any parameters (e.g., loss weights) that need to be modified? Could you provide original model weights of the paper data?