search

SungJaeShin / Gaussian_Splatting_Realtime_Render

3D Gaussian Splatting Real-time Rendered Image
6 stars 1 forks source link

fished #1

Closed huzhanxiaoxue closed 3 months ago

huzhanxiaoxue commented 3 months ago

hello，is it fished?

SungJaeShin commented 3 months ago

hello，is it fished?

Hello!

The trained Gaussian splatting model shown in the results was trained using a dataset of pinhole images and based on COLMAP.

For additional information, the data was acquired with a RealSense D455 sensor at a resolution of 640 x 480. The intrinsic parameters were obtained using Kalibr.

So if you want to try it yourself, follow all the methods I've mentioned, and it should work.

Thank you! :)

huzhanxiaoxue commented 3 months ago

hello，is it fished?

Hello!

The trained Gaussian splatting model shown in the results was trained using a dataset of pinhole images and based on COLMAP.

For additional information, the data was acquired with a RealSense D455 sensor at a resolution of 640 x 480. The intrinsic parameters were obtained using Kalibr.

So if you want to try it yourself, follow all the methods I've mentioned, and it should work.

Thank you! :)

Thank you for your answer, but I don't understand why COLMAP is used for training. Isn't VINS used here for pose and point cloud input? In the command python3 realtime_render.py -m -s , is the TRAINED_MODEL trained using COLMAP? Does this mean we first train on the scene, and then use VINS for mapping? I'm wondering if it's possible to combine VINS and 3DGS for real-time SLAM mapping?

SungJaeShin commented 3 months ago

hello，is it fished?

Hello! The trained Gaussian splatting model shown in the results was trained using a dataset of pinhole images and based on COLMAP. For additional information, the data was acquired with a RealSense D455 sensor at a resolution of 640 x 480. The intrinsic parameters were obtained using Kalibr. So if you want to try it yourself, follow all the methods I've mentioned, and it should work. Thank you! :)

Thank you for your answer, but I don't understand why COLMAP is used for training. Isn't VINS used here for pose and point cloud input? In the command python3 realtime_render.py -m -s , is the TRAINED_MODEL trained using COLMAP? Does this mean we first train on the scene, and then use VINS for mapping? I'm wondering if it's possible to combine VINS and 3DGS for real-time SLAM mapping?

Yes, that's correct. The trained model uses COLMAP for training.

This algorithm focuses on real-time rendering by taking the pose output from VINS as input and using the pre-trained GS Map. Therefore, COLMAP was necessary for creating the pre-trained GS Map.

VINS publishes local poses and odometry, which are then received by the realtime_render.py script to show real-time rendering for those poses. While it is possible to combine VINS and 3DGS for real-time mapping (like SLAM), this algorithm does not aim to achieve that goal.

huzhanxiaoxue commented 3 months ago

OK I know that.Thank you