blender的最大优点是可以用python进行交互。

经过一番google，我写了一组脚本，输出每一帧中无人机的屏幕坐标，作为训练用的ground truth.

get_obj_coord.py

<code class="language-python"># small helper script for CV+ML trainning ground truth data generation.
# by Qin Yongliang

import bpy
import bpy_extras

# get coordinate of object relative to screen.
def get_obj_coord(object_name):
    target = bpy.data.objects[object_name]
    
    # get object's location relative to world.
    target_loc = target.matrix_world.to_translation()
    
    cam = bpy.data.objects['Camera']
    
    scene = bpy.context.scene

    # relative to screen, where (0,0) represents left bottom of camera frame.
    co2d = bpy_extras.object_utils.world_to_camera_view(scene,cam,target_loc)

    return co2d

# get coordinate of object named 'Drone'.
def get_drone_coord():
    c = get_obj_coord('Drone')
    print('drone at screen coord:',c)
    return c
</code>

然后运行下面的脚本：

<code class="language-python">import get_obj_coord as g
import bpy

scene = bpy.context.scene
fs = scene.frame_start
fe = scene.frame_end

def get_drone_coord_list(startframe,endframe):
    list = []
    for i in range(startframe,endframe):
        scene.frame_set(i)
        c = g.get_drone_coord()
        list.append(c)
    return list
        
file = open("c:/coords.csv", "w")
list = get_drone_coord_list(fs,fe)
file.write('x,y,z\n')
for c in list:
    file.write('{},{},{}\n'.format(c.x,c.y,c.z))

file.close()
</code>

效果：

其中x和y是屏幕坐标，(0,0)是取景框左下角，(1,1)是取景框右上角。z是从相机到目标的距离。