和我的想法非常相近，完成度也很高啊喂。可以参考他的实现思路，引用的文章等等。

Intro

数字孪生(DT)作为现实世界非常精确的映射虽然可以用于高精度的训练但是生产DT资产过于繁琐且没有泛化性，不能做到zero-shot。
数字表亲(DC)通过比对模型特征，从模型库中选择类似的表亲模型，用于重建场景训练机械臂。让机械臂针对不同第一次见的场景具有泛化性。
（a）它减少了手动微调的需要，以保证一定的保真度，从而能够完全自动化地创建数字表亲，（b）它通过提供一组增强的场景来训练机器人策略，从而有助于更好地应对原始场景中的变化。

Methodology

ACDC is our automated pipeline for generating fully interactive simulated scenes from a single RGB image, and is broken down into three steps:
(1) an extraction step, in which relevant object masks are extracted from the raw input image
(2) a matching step, in which we select digital cousins for individual objects extracted from the original scene
(3) a generation step, in which the selected digital cousins are post-processed and compiled together to form a fully-interactive, physically-plausible digital cousin scene.

Goal

Estimate accurate 6D poses of multiple known objects in a 3D scene captured by multiple cameras with unknown positions

Challenges

• object pose hypotheses made in individual images cannot easily be expressed in a common reference frame when the relative transformations between the cameras are unknown(相机相对位置未知)
• the single-view 6D object pose hypotheses have gross errors in the form of false positive and missed detections（由于视角遮蔽，会存在误报和错漏的情况）
• the candidate 6D object poses estimated from input images are noisy as they suffer from depth ambiguities inherent to single view methods.（深度信息通常没那么精准）

Approach

自定义 Gen6d 物体

仓库：https://github.com/liuyuan-pal/Gen6D

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

步骤指令：

1
2
3
4
5
6
7
8
9

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

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

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

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

可以看到效果很好：

6-D Pose Estimation Survey

Model based (CAD model)

State of The Art: Foundation Pose (https://github.com/NVlabs/FoundationPose)

RGB

CASAPose (https://github.com/fraunhoferhhi/casapose?tab=readme-ov-file)

MegaPose (https://github.com/megapose6d/megapose6d)

D-RGB

MegaPose (https://github.com/megapose6d/megapose6d)

OVE6D (https://github.com/dingdingcai/OVE6D-pose)

Non-model

Gen6D (https://github.com/liuyuan-pal/Gen6D)

Yolov5 (https://github.com/cviviers/YOLOv5-6D-Pose)