Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

PMF-ResNet50 on SemanticKITTI Validation Set? #6

Closed
haibo-qiu opened this issue Dec 24, 2021 · 4 comments
Closed

PMF-ResNet50 on SemanticKITTI Validation Set? #6

haibo-qiu opened this issue Dec 24, 2021 · 4 comments

Comments

@haibo-qiu
Copy link

Hi~

Thanks for the open-source repo of your excellent work!

I notice that PMF-ResNet50 significantly outperforms PMF-ResNet34 on nuScenes Validation Set, and you even adopt ResNet101 on SensatUrban Test Set.

However, the result of PMF-ResNet50 (or deeper backbone) on SemanticKITTI Validation Set is unavailable. Did you try it before? Intuitively, it will also bring gains. Or did I miss something important?

@ICEORY
Copy link
Owner

ICEORY commented Dec 24, 2021

Unfortunately, using ResNet-50/101/152 on SemanticKITTI did not bring improvements to the model performance as we only use a limited subset (front-view only) of SemanticKITTI for training.

Recently, we found that using conformer as the backbone of the image stream can bring a 2.2% improvement on SemanticKITTI.

You can try more backbones to improve the performance~

@ICEORY
Copy link
Owner

ICEORY commented Dec 24, 2021

The model of conformer is pre-trained on ImageNet-21k and then fine-tuned on ImageNet-1k.

@haibo-qiu
Copy link
Author

Thanks a lot!

As for your mentioned conformer, I will give it a try :-)

@myself0816
Copy link

不幸的是,在SemanticKITTI上使用ResNet-50 / 101 / 152并没有提高模型性能,因为我们只使用SemanticKITTI的有限子集(仅前视图)进行训练。

最近,我们发现使用构象作为图像流的主干可以带来2.2%的SemanticKITTI改进。

可以尝试更多骨干网提升性能~

你好,我尝试使用conformer代替原本的resnet34,并且也加载了预训练权重,目前看到一影像流的分割精度很低,比resnet要差,请问您当时是怎样使用conformer的,有什么是我遗漏的吗?希望可以您进一步交流,这是我的qq1254952619

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

3 participants