search

DekuLiuTesla / CityGaussian

[ECCV2024] CityGaussian: Real-time High-quality Large-Scale Scene Rendering with Gaussians
https://dekuliutesla.github.io/citygs/
Other
329 stars 18 forks source link

did you used raw resolution photos? #14

Closed smart4654154 closed 1 month ago

smart4654154 commented 1 month ago

I have read your work and it is great. I have some confusion. Could you please answer it I am also using data from matrix city (small city), 1.May I ask if you used Colmap to match all 7600 pieces of data at once?

  1. May I ask if you trained coarse Gaussian using raw resolution photos (1920 * 1080) as input?

For question 1, I have reviewed the colmap results you provided and found that the file is quite large. I guess the answer is that colmap matched all 7600 original resolution data at once For question 2, I find "resolution : -1"in mc_aerial_coarse.yaml. I guess the answer is that you trained coarse Gaussian using the original resolution photo as input

If I guessed wrong, please correct me. Based on my guess, I think this seems to be a difficult task Colmap matches all 7600 pieces of data at once, which takes a long time and may cause the computer to crash 7600 original resolution photos (1920 * 1080) should be out of memory. Looking forward to you answering my questions Thank you very much.

DekuLiuTesla commented 1 month ago

Good Question! Here is the answer:

  1. We use the ground-truth poses offered by MatrixCity dataset and separately match the train and test sets. And this will be faster and more robust than match from scratch. But indeed it still costs a lot of time.
  2. By applying resolution=-1, the image size is sampled to 1600x900, as default behavior in 3DGS. In addition, we apply CacheDataloader from Gaussian Lightning, the memory cost becomes acceptable, while the fine-tuning of each block can be finished within around 40 minutes on A100