Ruopeng Gao

icon indicating a website link www.ruopenggao.com

icon company Nanjing University icon location Nanjing, China icon location Ph.D student, @MCG-NJU, Nanjing University. 世界上大部分事，都没有太大意义，真理与爱除外。

Results 100 comments of Ruopeng Gao

图像生成器（Generator）中的数据分布信息从何而来？

> 你好，这段话针对的是公式（1）在有原始数据下的情况。 > > 在没有数据的情况下，也就是p(x)未知时我们采用随机向量z生成数据，见公式(5)。 您说的第一点确实是我的问题，我之前未能注意这两个公式的区别。 关于第二点，我想请问一下这里的 p_z 描述为一个分布，z 从 p_z 中取随机分布，请问这里的分布 p_z 是否只是单纯表示一个规定了 w, h, c，而每一个像素可以是任意颜色的图像分布？而不是对原始数据集进行了某些方面的统计从而得到的更加特别的分布 p_z？ 感谢您的回复，再次打扰。

About the confused relationship between DetA and predict IDs.

Yes, in my understanding, I also think that the ID alignment result will not affect the `DetA` result. But in the [paper](https://link.springer.com/article/10.1007/s11263-020-01375-2), they explained in page 555: > **Matching to...

EvalAI submission error!

I have a similar issue [here](https://github.com/bdd100k/bdd100k/discussions/327).

about training time

They are all trained on 8 GPUs.

Invite researchers in the area of MOT to join in this repo

Hi, I am willing to participate in the maintenance of this repository. Email: [email protected] GitHub ID: HELLORPG

Invite researchers in the area of MOT to join in this repo

OK. Thanks a lot. 😄

When I trained MOTR using my own dataset (with only one class), the assert error (boxes1[:, 2:] >= boxes1[:, :2]).all() occurred during the training process, with all values of boxes being NaN. I tried to lower the learning rate, from lr=2e-4 to lr=2e-6, and lr_backbone from 2e-5 to 2e-7. It did not help, and I did not change the code.

After how many iterations will this problem occur?

请教关于训练时使用的显存大小与训练效果的关系讨论

一般来说，训练的效果和 batch size 是具有很大的相关性的，所以复现的话一般需要同样的 batch size 才能复现。

用两个A100训练，只能达到65 Training with two A100s can only reach 65

原文采用的 batch size 应该是 8，如果使用两张 GPU 的话应该是无法得到同样的 batch size 的，这确实会对最终的结果产生影响。

visibility

If you are talking about the training annotation, I don't think so. The conf scores in the training gts are all set to 1.0, just like DanceTrack. But the visibility...