hrzhou2 / AdaptConv-master

Adaptive Graph Convolution for Point Cloud Analysis
110 stars 12 forks source link

Classification performance #2

Open PointCloudYC opened 3 years ago

PointCloudYC commented 3 years ago

Hi! Thanks for sharing your enlightening work.

I have a question for ModelNet40 classification accuracy. Without changing any code or settings, I successfully run your code on my machine (1 Nvidia 3090 GPU and Ubuntu18.04). But the best accuracy I can obtain is 92.8% which is 0.8% lower than the number reported in your paper.

Can you help point out the possible factors influencing the final performance on the ModelNet40?

Thanks.

image

hrzhou2 commented 3 years ago

So, first of all, make sure the dataset was correct.

From my experience, there may be some randomness in the classification scores. I think you should just run a few more times and see what you get.

If necessary, pretrained models can be released in some future updates.

PointCloudYC commented 3 years ago

So, first of all, make sure the dataset was correct.

From my experience, there may be some randomness in the classification scores. I think you should just run a few more times and see what you get.

If necessary, pretrained models can be released in some future updates.

Thanks for your prompt reply. I hyper-tuned the model(k=10, use sgd), the best performance I obtain reach 93.47% OA. image