search

Haiyang-W / UniTR

[ICCV2023] Official Implementation of "UniTR: A Unified and Efficient Multi-Modal Transformer for Bird’s-Eye-View Representation"
https://arxiv.org/abs/2308.07732
Apache License 2.0
276 stars 16 forks source link

What's the performance on the nuScene without pre-trained on ImageNet and nuImage ? #26

Closed gwenzhang closed 1 month ago

gwenzhang commented 1 month ago

Hi, thank you for your excellent work! Can you provide the performance of UniTR on the nuscene without the pre-trained parameters?

nnnth commented 1 month ago

We haven't tried training UniTR without pre-trained parameters because using pre-trained parameters has been very common in previous work. Typically, not using pre-training lowers performance. You can experiment by adjusting the command parameters to see how it performs. If you have any more questions or need further assistance, feel free to ask.

gwenzhang commented 1 month ago

Would it be possible to provide the configuration files for pretraining on nuImage and ImageNet? Alternatively, could you share which open-source framework this is based on? Any information or resources would be greatly appreciated. Thanks in advance for your help

We haven't tried training UniTR without pre-trained parameters because using pre-trained parameters has been very common in previous work. Typically, not using pre-training lowers performance. You can experiment by adjusting the command parameters to see how it performs. If you have any more questions or need further assistance, feel free to ask.

nnnth commented 1 month ago

Unfortunately, the original configuration files we used for pretraining are lost. We conducted ImageNet pretraining using the MMPretrain and nuImage pretraining using MMDetection3D. For nuImage pretraining, you can refer to the configuration file at this link: nuImage configuration.

Both frameworks have mature configurations that you can adapt for your models and use for pretraining.

gwenzhang commented 1 month ago

Unfortunately, the original configuration files we used for pretraining are lost. We conducted ImageNet pretraining using the MMPretrain and nuImage pretraining using MMDetection3D. For nuImage pretraining, you can refer to the configuration file at this link: nuImage configuration.

Both frameworks have mature configurations that you can adapt for your models and use for pretraining. Thank you so much for your help!