Hello, I trained cifar10-lt and cifar100-lt with the default parameters in the code, but the results I obtained are quite different from those in the paper.
Could you please tell me how can I achieve the results in the paper? Could you provide your commands, or are there any other parameters or settings that need to be noted?
This is what I used for training：

• train：python train.py --gpu 0 --ds cifar100 --Lambda1 0.05 --Lambda2 0.05 --Lambda3 0.1 --drp ../long-tailed-ood-detection/SCOOD_dataset/data/images --srp ./checkpoints
• test：'python test.py --gpu 0 --ds cifar100 --dout shvn --drp ../long-tailed-ood-detection/SCOOD_dataset/data/images --ckpt_path ./checkpoints/cifar100-0.01-OOD300000/ResNet18/e100-b128-256-adam-lr0.001-wd0.0005_Lambda10.05-Lambda20.05-Lambda30.1/replay3'
  and this is the result i got：
  

Hi, thank you for your great work~ I have a few questions regarding my understanding of the paper:
1.According to Figure 1, I understand that auxiliary OOD data are also used during training, similar to outlier exposure, is that correct?
2.In the "Debiased Large Margin Learning Calibration" module, the schematic diagram does not illustrate the operation of pulling OOD categories. However, the text mentions "since we have already pulled OOD samples together in the joint LTR and outlier class learning in Eq. 2", does this mean there is an operation to pull the entire OOD category?
If so, considering that OOD data have diverse categories and their distribution is theoretically unknown, how can we ensure they cluster together in the feature space?

I would greatly appreciate it if you could provide some clarification.