NJU-3DV / Relightable3DGaussian

[ECCV2024] Relightable 3D Gaussian: Real-time Point Cloud Relighting with BRDF Decomposition and Ray Tracing
https://nju-3dv.github.io/projects/Relightable3DGaussian/
Other
458 stars 34 forks source link

The training time at stage 2 using our own dataset was too long #41

Open 2019213039 opened 2 months ago

2019213039 commented 2 months ago

First of all, I really want to thank you for your work. It's a great job!

The problem I ran into was the time it took me to train stage2 using my own dataset (converted to DTU dataset format). My own dataset consists of 67 views. My environment configuration is the same as your recommended environment. The device used is an RTX4060 with 8G video memory. My model had a total of 227,970 3D Gaussians after completing the first stage of training. Your paper says that the second phase of training should be very rapid:

Speed and Memory To train a NeRF synthetic scene, our model typically requires approximately 16 minutes and 10 GB of memory. Notably, the visibility baking in the second stage is remarkably brief, lasting only a few seconds. Following this optimization, our model attains real-time rendering across different lighting conditions, achieving 120 frames per second.

Is the training time too long because my GPU memory is too low? If yes, what can I do to speed up the second phase of training?

Please let me know if you need more information from me. My wechat ID is lyc1600378878