agent的选择

[mengmeng6660]

对于这种零样本问题 Zero-Shot Object Goal Visual Navigation 使用navigation_agent 而您这使用semantic是否合理

[rginjapan]

@wanhaozhangxiaorui 我也注意到了这个问题，想问下一下这两种agent模式有什么区别？？

[mengmeng6660]

@rginjapan 我看Zero-Shot Object Goal Visual Navigation和TDANet: Target-Directed Attention Network For Object-Goal Visual Navigation With Zero-Shot Ability这两篇论文都提到了 零样本导航 是训练时只用到可见类 在获取当前图像的物体检测框时会将不可见类排除 而这篇论文在训练时也会检测到不可见类 只不过不会当做目标对象罢了 这种改变应该不算零样本导航了吧 毕竟训练时也见过了 只不过没当做目标罢了 我是这么理解的 有可能理解错误 还望指正

[rginjapan]

@wanhaozhangxiaorui 嗯嗯。跟我想的是一样的。 所以我之前一直不理解他为什么能把结果从20多提升到60多，毕竟他的State matrix也没什么大的不同，GCN也是。还有TDANet的提升也很大，不过没有代码可以测。。。我现在在train this work on NavigationAgent, 看看效果和ZSON有多大提升。。。

[rginjapan]

@wanhaozhangxiaorui 你可以试试跑之江的ICRA 2024 ：https://github.com/nuoxu/AKGVP 里面也有zero-shot的evaluation，我也有很多困惑。。。

[mengmeng6660]

我看过那篇论文 和这种基于相似度的有点区别 等我跑跑代码 有收获我就跟你交流

[rginjapan]

@wanhaozhangxiaorui I evaluated the self-trained model under NavigationAgent class, the SR on unseen is only 0.09, with 0.606 on seen classes, which are inferior than ZSON.

[mengmeng6660]

I don't know the details of your model, but from your performance metrics, it seems that navigational performance is strongly correlated with target features.

[rginjapan]

It is the same model in this repo, I just used the NavigationAgent class instead of the SemanticAgent for training and evaluation.

[mengmeng6660]

@rginjapan I understand your point, but there seems to be a misunderstanding. For zero-shot tasks, the zero-shot objects should be excluded during training, not included. Therefore, during training, we use "navigation," and during testing, we switch to using "semantic

[rginjapan]

我用semanticagent跑了一遍（训练时navigationagent），unseen: 0.616, seen, 0.579. 那么Zero-Shot Object Goal Visual Navigation这个好像训练和测试都是用的navigationagent啊？

[chuangwaipiaoluohuaxiang]

我用semanticagent跑了一遍（训练时navigationagent），unseen: 0.616, seen, 0.579. 那么Zero-Shot Object Goal Visual Navigation这个好像训练和测试都是用的navigationagent啊？

你好，请问你可以跑通CIRN的代码吗，我遇到了一些问题。