Posted 2024-08-01Updated 2024-09-28Notea minute read (About 169 words)

手册：https://github.com/liuyuan-pal/Gen6D/blob/main/custom_object.md

步骤指令：

python prepare.py --action video2image --input data/custom/part1/ref.mp4 --output data/custom/part1/images --frame_inter 10 --image_size 960  

python prepare.py --action sfm --database_name custom/part1 --colmap D:\COLMAP\COLMAP-3.7-windows-cuda\COLMAP.bat        

# do some cloudcompare thing

python predict.py --cfg configs/gen6d_pretrain.yaml --database custom/part1 --video data/custom/part1/test.mp4 --resolution 460 --output data/custom/part1/test --ffmpeg ffmpeg

ffmpeg -framerate 30 -i %d-bbox.jpg -c:v libx264 -r 30 -pix_fmt yuv420p output.mp4

关于判定不准确怎么解决：https://github.com/liuyuan-pal/Gen6D/issues/29

Posted 2024-08-01Updated 2024-09-28Notea minute read (About 129 words)

6d pose -- unity coordinate

6d pose -> unity coordinate

unity 使用左手坐标系，普遍的 6d 算法使用右手坐标系，所以得出[R;t]后需要做一步针对 y 轴的反射变换

def right_to_left_hand_pose_R(R):
    # 定义反射矩阵
    M = torch.tensor([
        [ 1,  0,  0],
        [ 0, -1,  0],
        [ 0,  0,  1]
    ], dtype=R.dtype)
    
    # 转换旋转矩阵
    R_prime = M @ R @ M
    
    return R_prime

def right_to_left_hand_pose_t(t):
    # 定义反射矩阵
    M = torch.tensor([
        [ 1,  0,  0],
        [ 0, -1,  0],
        [ 0,  0,  1]
    ], dtype=t.dtype)

    # 转换位移向量
    t_prime = M @ t

    return t_prime