VFA

VFA copied to clipboard

Hello. Thanks for your great work! In your paper Meta R-CNN++ was presented as a stonger baseline. But it seems that Meta R-CNN++ is similar to the Meta R-CNN implementation in mmfewshot. I would like to ask whether they are the same.

Thanks! They are similar in parameter freeze, but different in cosine classifier and last layer weight init. Please check Tab.1 in our paper.

Thank you for your timely reply. Do you use this setting in your experiments? I took a look at the code and found that both Meta R-CNN++ and mmfewshot's implementation use linear classier and randomly init the last layer.

Thanks! They are similar in parameter freeze, but different in cosine classifier and last layer weight init. Please check Tab.1 in our paper.

Could you please provide me with the t-SNE visualisation code on VFA? Thank you very much and I would be happy to talk to you.

@hhxzzy Actually, we do not use cosine classifier and last layer init in the open-sourced version as it takes additional checkpoint combination step, similar to https://github.com/ucbdrive/few-shot-object-detection/blob/master/tools/ckpt_surgery.py If we want to report the GFSOD results, however, it is neccesary to use Meta R-CNN++ to maintain accuracy on base classes.

@873552584 We do not use t-SNE in the paper. Do you mean Similarity matrix visualization? You can store the latent region features on disk and calculate their cosine similarities. Then we use Origin software to plot the similarity matrix.

Do you mean that the open-sourced version is for FSOD task, which dosen't use cosine classifier and last layer copy init, and the version for GFSOD uses cosine classifier and copy init? And I have a question about the GFSOD result. Is the GFSOD result calculated as the average of all base and novel classes?

@hhxzzy We do not report bAP here, but the full GFSOD results should contain AP, bAP and nAP.

So the results in table7 is nAP under the setting of GFSOD?

Yes. But we only report nAP.

Is the open-sourced code under the setting of FSOD, which doesn't use cosine classifier and copy init, while the GFSOD setting uses cosine classifier and copy init?

@873552584 我们在论文中不使用 t-SNE。 你的意思是相似矩阵可视化？您可以将潜在区域特征存储在磁盘上并计算它们的余弦相似度。然后我们使用Origin软件绘制相似度矩阵。

Thanks

@hhxzzy Yes. This difference makes minor effect on nAP.

Is the base training checkpoint provided in model zoo under the setting of FSOD or GFSOD?

@873552584我们在论文中不使用 t-SNE。 你的意思是相似矩阵可视化？ 您可以将潜在区域特征存储在磁盘上并计算它们的余弦相似度。 然后我们使用 Origin绘制相似矩阵的软件。

I mean t-SNE to visualize such a clustering or classification graph.

@873552584我们在论文中不使用 t-SNE。 你的意思是相似矩阵可视化？ 您可以将潜在区域特征存储在磁盘上并计算它们的余弦相似度。 然后我们使用 Origin绘制相似矩阵的软件。

I mean t-SNE to visualize such a clustering or classification graph.

您好，请问您解决这个问题了吗？