Why in custom data, the accuracy of Detectron 2 is 22% higher than MIC model?

[behnaz-sadeghigol]

I have a custom data. With the MIC model, I got 22% accuracy, while I ran my custom data with 300 iterations in the Fester-RCNN (Detectron 2) model, so that the training and validation data were from custom synthetic data, and the test data was from real custom data. Its accuracy was 44%. What do you think could be the reason? I defined the data in the MIC model as follows:

"cityscapes_fine_instanceonly_seg_train_cocostyle": { "img_dir": "/home/sadeghigol/sadeghi/testorch2/dataset/syn_img/img_dir/train", "ann_file": "/home/sadeghigol/sadeghi/testorch2/dataset/syn_img/ann_dir/train/jltv_custom_synthetic_train.json" }, "cityscapes_fine_instanceonly_seg_val_cocostyle": { "img_dir": "/home/sadeghigol/sadeghi/testorch2/dataset/syn_img/img_dir/val", "ann_file": "/home/sadeghigol/sadeghi/testorch2/dataset/syn_img/ann_dir/val/jltv_custom_synthetic_val.json" }, "cityscapes_fine_instanceonly_seg_test_cocostyle": { "img_dir": "/home/sadeghigol/sadeghi/testorch2/dataset/syn_img/img_dir/test", "ann_file": "/home/sadeghigol/sadeghi/testorch2/dataset/syn_img/ann_dir/test/jltv_custom_synthetic_test.json" }, "foggy_cityscapes_fine_instanceonly_seg_train_cocostyle": { "img_dir": "/home/sadeghigol/sadeghi/testorch2/dataset/real_img/img_dir/train", "ann_file": "/home/sadeghigol/sadeghi/testorch2/dataset/real_img/ann_dir/train/modified_coco_labels_train.json" }, "foggy_cityscapes_fine_instanceonly_seg_val_cocostyle": { "img_dir": "/home/sadeghigol/sadeghi/testorch2/dataset/real_img/img_dir/val", "ann_file": "/home/sadeghigol/sadeghi/testorch2/dataset/real_img/ann_dir/val/jltv_custom_val.json" },