Test the multi-view camera pose generation.
The camera pose of the input image of Zero123++ (data/edited_chair.png) is:
"rotation": 0.012566370614359171, "transform_matrix": [ [ 0.3260957896709442, 0.14048941433429718, -0.934839129447937, -3.7684571743011475 ], [ -0.9453368186950684, 0.04846210405230522, -0.3224746286869049, -1.2999367713928223 ], [ 0.0, 0.9888953566551208, 0.1486130952835083, 0.5990785360336304 ], [ 0.0, 0.0, 0.0, 1.0 ] ]
Six-view images generated by Zero123++:
data/chair_6views/train
Their elevation and azimuth are shown in Zero123++ paper.
Conda env: use the same packages as gaussian-splatting.
Put the pretrained chair ply file in this folder:
output/chair_initial/point_cloud/iteration_30000
Camera pose file:
data/chair_6views/transforms_train.json
To validate the correction of camera pose, finetune pre-trained gaussian:
python train.py --iterations 5000 -s data/chair_6views -m output/chair_initial --finetuning True
The output will be in this folder: output/chair_initial/finetuned
To render the finetuned gaussian, first move the finetuned model folder (output/chair_initial/finetuned/iteration_5000) from output/chair_initial/finetuned to output/chair_initial/output, then run:
python render.py -m output/lego_initial --iteration 5000
To render the initial gaussian, run:
python render.py -m output/lego_initial --iteration 30000
Then the render images will be in:
output/chair_initial/train/ours_5000/renders