Bo-Wen Yin

Thanks for your attention to our work! It is suggested that use the fixed random seed in training process and dataloader. You can change line 27 in 'train.sh' from '--no-use_seed'...

We upload the training/test.txt of kitti-360 in Baidu Netdisk: https://pan.baidu.com/s/1-CEL88wM5DYOFHOVjzRRhA?pwd=ij7q

您好。您说的是自己的数据集可以训练和测试?还是说只有一些想要直接测试的图片。前者的话可以参考这个[link](https://github.com/VCIP-RGBD/DFormer/tree/main/figs/application_new_dataset)。后者的话如果是想用NYU或者SUNRGBD的label来得到对应的分割图，可以用我们训练好的pth，把你的图片当成测试集来使用。

> 你好，这个模型可以用自己的数据集进行训练吗？在用自己的数据做测试？ 可以的，你可以参考这个教程来添加您自己的数据集到这个框架下[apply on your own data](https://github.com/VCIP-RGBD/DFormer/tree/main/figs/application_new_dataset)

> 你好，我想问一下，自己的电力数据集，只有原图，没有深度图，那用这个模型是不是不能生成深度图，并且利用二者进行分割？感谢！ > […](#) 我这个项目下的模型是RGB-D输入的，只有RGB图可能不太合适，但是你可以使用[depthpro](https://www.bing.com/ck/a?!&&p=a54b0c4ff1553c6a10621b158a1b3db74d8a34aaf74bdf798eb43dea957c454aJmltdHM9MTc1MjM2NDgwMA&ptn=3&ver=2&hsh=4&fclid=2916e694-be24-65fc-38f3-f336bff66481&u=a1aHR0cHM6Ly9naXRodWIuY29tL2FwcGxlL21sLWRlcHRoLXBybw&ntb=1)或[depthanything](https://www.bing.com/ck/a?!&&p=d7b279baf27a5771ecea11b680ec2e00d9d0b80d19f4cf3e51b779b9aa0a6aacJmltdHM9MTc1MjM2NDgwMA&ptn=3&ver=2&hsh=4&fclid=2916e694-be24-65fc-38f3-f336bff66481&u=a1aHR0cHM6Ly9naXRodWIuY29tL0xpaGVZb3VuZy9EZXB0aC1Bbnl0aGluZw&ntb=1)来生成深度图使用，现在这些深度估计方法效果已经很好了。

Thank you for your interest in our work! The DFormerv2-S model was fine-tuned using two NVIDIA 3090 GPUs, while larger model scales required four 3090 GPUs for optimization. Fine-tuning durations...

> 你好，我想问一下验证的时候显存溢出了，用的两张5090，训练集就1000张，这是正常的嘛 验证时显存溢出，应该检查验证集图片的尺寸和验证时的batch size, val_batch_size[this line](https://github.com/VCIP-RGBD/DFormer/blob/1b1cf2ee24624cf8fd23908b5cd94ba1eaef197c/utils/eval.py#L65C5-L70C44),图片较大的话可以考虑调小val_batch_size

We encountered some issues retrieving the previous model from the server. We have retrained a DFormer-S pretrained solely on depth of ImageNet for you, which can achieve better results on...

Hi, thanks for attention to our work! You can train our pre trained model in unsupervised RGBD semantic segmentation scenarios. I think our RGBD pre training and specially designed models...

Thanks for your attention to our work! If you are finetuning our trained weight with a new dataset, we recommend to keep consistent with our pretraining([code](https://github.com/VCIP-RGBD/RGBD-Pretrain/blob/12947c1ca521126fa882ad5e1859ac9333b8ac4b/data/transforms_factory.py#L120)): ` transforms.Normalize( mean=torch.tensor(mean), std=torch.tensor(std))`...