search

huixiancheng / CENet

[ICME 2022] CENet: Toward Concise and Efficient LiDAR Semantic Segmentation for Autonomous Driving
MIT License
99 stars 13 forks source link

About pretrained model #7

Open ohosh opened 2 years ago

ohosh commented 2 years ago

Dear author,

Thanks for the sharing code.

when I use the pretrained model about Kitti result you gave to run the datasets, the results can not reach the effect in your article. I don't know if there is a problem with the model I used, because there are multiple models in the file. I would like to ask which model you used to get the results in your article.

Thanks！

huixiancheng commented 2 years ago

Hi. The KITTI results only contain our training logs and results under 64x512 inputs. The model under the file named “512-594” you can get a test accuracy of 59.4. The file named “512+vaild-607” is the result after adding the validation set to the training and fine-tuning it, using this model test you should get a test set accuracy of 60.7 as reported in our paper. Very sorry, pre-trained models and logs under larger inputs may not be considered for release

cardwing commented 2 years ago

@ohosh @huixiancheng I have trained the CENet which can achieve 65 mIoU on SemanticKITTI test set with 64x512 input resolution. The trained model will be put in https://github.com/cardwing/Codes-for-PVKD.

cardwing commented 2 years ago

More powerful range-image-based models built on the awesome CENet will also be put in that repo (70+ mIoU on SemanticKITTI test set).

huixiancheng commented 2 years ago

@cardwing Incredible. :open_mouth: :scream: :see_no_evil: That's awesome and amazing work. I think this will drive further development of the range-based methods. :thumbsup::thumbsup::thumbsup: Looking forward to the release.

cardwing commented 2 years ago

The CENet with 64x512 input resolution has been uploaded to https://github.com/cardwing/Codes-for-PVKD. The reproduced performance (67.6% mIoU) is much higher than the reported value on SemanticKITTI test set.

huixiancheng commented 2 years ago

Glad to hear this. :clap: :thumbsup: Modify README to point to that great work and Repo. :point_right:

huishuai13 commented 2 years ago

The CENet with 64x512 input resolution has been uploaded to https://github.com/cardwing/Codes-for-PVKD. The reproduced performance (67.6% mIoU) is much higher than the reported value on SemanticKITTI test set.

Is the CENet model with 64x512 in your repository trained with distillation?