Can I ask if the weight you provided resnext101_DCN_160e_epoch_150.pth
was trained according to the steps in the paper on the Vietnamese dataset, particularly with maps? When I displayed the detection results, they were quite poor (even though I tested on the dataset you provided).