leo-frank / colmap-free-3d-gs

A PyTorch reproduction of colmap-free 3D Gaussian Splatting (https://arxiv.org/abs/2312.07504). We want to implement it and make it better.

Colmap-Free-3D-GS

A PyTorch reproduction of colmap-free 3D Gaussian Splatting (https://arxiv.org/abs/2312.07504). Since there is no open source codes, we want to implement it and make it better.

Coming before 4.19 Apr.

Features Implemented

colmap-free-3d-gs's Issues

Release the code

Hello, may I ask how much longer until you can release the code?

Hello, have you managed to solve the colmap-free method? I encountered some issues when not using colmap as well.

Hello, I encountered a failure in reconstructing shirtv1 from the speed+ dataset. I first converted the camera.json file and the camera parameters roe2.json file from the dataset into the format required by the nerf dataset, as shown below.
{
"camera_angle_x": 5.86e-06,
"rotation": 3.1415557843472173,
"frames": [
{
"file_path": "img000001.jpg",
"rotation": 3.1415557843472173,
"transform_matrix": [
[
1.0,
-1.0206827928449808e-16,
8.85350556627883e-18,
0.04852291750863577
],
[
2.943923360032076e-17,
0.36904495838715945,
0.9294115443058688,
-0.31807457560899577
],
[
-9.813077866773592e-17,
-0.9294115443058688,
0.36904495838715945,
7.86837091563897
],
[
0.0,
0.0,
0.0,
1.0
]
]
},
However, when I attempted the reconstruction using these parameters, the results were very poor, to the extent that there was no visible point cloud. Can someone assist me in resolving this issue? Thank you very much!
Below is the specific content of the camera.json file:
{
"Nu": 1920,
"Nv": 1200,
"ppx": 5.86E-6,
"ppy": 5.86E-6,
"fx": 0.017697236438454025,
"fy": 0.017697236438454025,
"ccx": 960,
"ccy": 600,
"cameraMatrix": [
[
3020.0062181662161,
0,
960
],
[
0,
3020.0062181662161,
600
],
[
0,
0,
1
]
],
"distCoeffs": [
-0.21249040440358169,
0.4443683447224509,
-0.00038703301037756828,
-0.00044885973454884538,
0.56835118403785934
]
}
Below is the specific content of the roe2.json file:
[
{
"filename": "img000001.jpg",
"q_vbs2tango_true": [
-0.82735873669985494,
0.56167385626039268,
-3.2327054604134184E-17,
-3.9737149996557807E-17
],
"r_Vo2To_vbs_true": [
0.048522917508635771,
-0.31807457560899577,
7.86837091563897
]
},
{
"filename": "img000002.jpg",
"q_vbs2tango_true": [
-0.82542466067382414,
0.56363503659809222,
-0.011931472870665516,
-0.029108676124239254
],
"r_Vo2To_vbs_true": [
0.051787825059848837,
-0.31742644621634647,
7.8671667361472455
]
},
{
"filename": "img000003.jpg",
"q_vbs2tango_true": [
-0.822602785207585,
0.56512824651642046,
-0.024000254897478319,
-0.05812667651651994
],
"r_Vo2To_vbs_true": [
0.055041116890462177,
-0.31667209568172283,
7.8658190773001042
]
},
Here is my final reconstruction result.
This is a paper on speed+.Adaptive Neural Network-based Unscented Kalman Filter for.pdf

[SIBR] -- INFOS --: Initialization of GLFW
[SIBR] -- INFOS --: OpenGL Version: 4.6.0 NVIDIA 551.61[major: 4, minor: 6]
[SIBR] -- INFOS --: Did not find specified input folder, loading from model path
[SIBR] ## ERROR ##: FILE C:\projects\gauss2\SIBR_viewers\src\core\scene\ParseData.cpp
LINE 560, FUNC sibr::ParseData::getParsedData
Cannot determine type of dataset at /D:\try_speed+\sunlamp
Number of input Images to read: 2511
Number of Cameras set up: 2511
[SIBR] -- INFOS --: Mesh contains: colors: 1, normals: 1, texcoords: 0
[SIBR] -- INFOS --: Mesh 'C:\Users\users\Downloads\result/input.ply successfully loaded. 1 meshes were loaded with a total of (0) faces and (100000) vertices detected. Init GL ...
[SIBR] -- INFOS --: Init GL mesh complete
[SIBR] -- INFOS --: Loading 299 Gaussian splats
[SIBR] -- INFOS --: Initializing Raycaster
[SIBR] -- INFOS --: Interactive camera using (0.009,1100) near/far planes.
Switched to trackball mode.