VCIP-RGBD
Congratulations on CVPR Acceptance! A Few Questions about DFormer-V2 Training and DELIVER Dataset Support
First of all, heartfelt congratulations on the acceptance of DFormer-V2 at CVPR! Your work is truly inspiring, and I have learned a lot from your previous V1 version. I have a couple of questions and would greatly appreciate your guidance: Training Setup: Was the DFormer-V2 model trained using two RTX 3090 GPUs as in the previous version? If possible, could you kindly share the hardware configuration and approximate training time for both pretraining and fine-tuning? DELIVER Dataset Support: I noticed that you’ve included evaluation results on the DELIVER dataset in your latest paper. May I ask if you have any plans to update the codebase to support the DELIVER dataset in the future? Thank you very much for your time and for your valuable contributions to the field. Wishing you continued success and even greater achievements in your research!
Thank you for your interest in our work!
The DFormerv2-S model was fine-tuned using two NVIDIA 3090 GPUs, while larger model scales required four 3090 GPUs for optimization. Fine-tuning durations spanned approximately 10, 15, and 20 hours for the respective model sizes, with pretraining phases taking about 80, 110, and 150 hours on eight 3090 GPUs. Regarding the Deliver dataset implementation, while the initial training utilized another framework, we are actively refactoring the code to ensure seamless integration with our main codebase. If you are in a hurry, you can also apply our model to the deliver framework. We will update the unified version once we finish and will notify you.
Wishing you success in your research endeavors!
