Switch camera pose from SE3 to quaternion, also support multiple camera poses for later dynamic object.

[wanmeihuali]

Switch camera pose from SE3 to quaternion, to simplify later pose optimization. Also, a simple idea is to assign an object id to each point, take some 6DOF pose estimation for moving objects as input, then handle rigid objects as points with different camera poses (movement of the object is transformed into the movement of the camera). In this PR, we first change the input format for rasterization. Will try to support the whole feature in later PR.

[github-actions[bot]]

Running experiment on sagemaker with git sha 9a30d168f59b32c154ea6f5b9abefd58344f2db6

[github-actions[bot]]

Training job camera-pose-qt-and-multi-obj-9a30d16-230629-080231-tat-truck created

[github-actions[bot]]

Training job camera-pose-qt-and-multi-obj-9a30d16-230629-080231-tat-truck completed.

Model url: s3://taichi-3d-gaussian-splatting-log/tat-truck/camera-pose-qt-and-multi-obj-9a30d16-230629-080231-tat-truck/output/model.tar.gz,

tensorboard output path: s3://taichi-3d-gaussian-splatting-log/tat-truck/camera-pose-qt-and-multi-obj-9a30d16-230629-080231-tat-truck/output/output.tar.gz

[github-actions[bot]]

Training job camera-pose-qt-and-multi-obj-9a30d16-230629-080231-tat-truck final metrics:

Latest Metrics

train:ssimloss	train:psnr	val:ssim	train:ssim	train:loss	train:num_valid_points	train:l1loss	val:loss	val:psnr	train:iteration
0.1402938961982727	26.190698623657227	0.8397595286369324	0.8597061038017273	0.05099417641758919	440195.0	0.028669243678450584	0.060268305242061615	24.510377883911133	30000.0

Max Metrics

val:5kssim	train:5kpsnr	train:7kssim	val:5kpsnr	train:psnr	val:ssim	train:ssim	train:5kssim	val:psnr	val:7kssim	val:7kpsnr	train:7kpsnr
0.773842990398407	22.222253799438477	0.7961171865463257	22.148059844970703	26.190698623657227	0.8397595286369324	0.8597061038017273	0.8031989932060242	24.510377883911133	0.7987189292907715	22.8942928314209	22.990148544311523