Problems encountered during training with DeepGlobe

[zts12]

Hello, I used DeepGlobe's 1024 images, cropped them to a size of 512, and then generated the required data using Sat2Graph method. However, I used the data shown in the picture for training
RNGDetPlusPlus， Using Python=main_stample. py -- savedir datatest1-- dataroot/data/-- imagesize 512-- ROI-SIZE 160
--Edge_mave_ahead_1ength 40-- num_queries 15-- noise 8-- max_num_frame 8000 and CUDA_VISIBLE-DEVICES=0,1,2,3 python-m torch. distributed-launch -- nproc_per_node=4 main_train.exe -- savedir RNGDet
--dataroot /data/ --batch_size 20 --ROI_SIZE 160 --nepochs 50 --multi_GPU --backbone resnet101 --eos_coef 0.2
--lr 1e-4 --lr_backbone 1e-4 --weight_decay 1e-5 --noise 8 --image_size 512
--candidate_filter_threshold 40 --logit_threshold 0.75 --extract_candidate_threshold 0.55 --alignment_distance 12
--The current dataset consists of 24904 images, with 19931 for training, 3732 for validation, and 1241 for testing. However, the results obtained so far are very poor, precision/recall/f1: 0.029/0.049/0.037， My contact information is [email protected] If it's convenient, you can contact me by email and I can provide more data on the training process. I am currently very confused about this and thank you very much for your help.