I will start with the issue in your code and present my solution. I'd recommend in the future to paste your code directly into the question, as that will make the question more approachable, without having to download your blend file.
For others, here's the code you had issues with:
import bpy
import bmesh
print("-----------------start---------------------")
cam = bpy.data.objects['camera']
cam_pos = cam.location
scene = bpy.context.scene
visible_objs = []
# loop all mesh obj
bpy.ops.object.mode_set(mode='OBJECT')
bpy.ops.object.select_all(action='DESELECT')
obj_cout = 0
for obj in scene.objects:
if obj.type == 'MESH' and obj.name not in ['ground']:
obj_cout += 1
print("At obj: ", obj_cout, " ", obj.name)
# reset
bpy.ops.object.mode_set(mode='OBJECT')
bpy.ops.object.select_all(action='DESELECT')
bpy.context.view_layer.objects.active = None
# set active
bpy.context.view_layer.objects.active = obj
# go to edith mode
bpy.ops.object.mode_set(mode='EDIT')
obj = bpy.data.objects[obj.name]
mesh = obj.data
bm = bmesh.from_edit_mesh(mesh)
selected_verts = 0
all_verts = 0
for v in bm.verts:
all_verts += 1
dir = cam_pos - v.co
dir.normalize()
hit, loc, normal, index, ob, mat = bpy.context.scene.ray_cast(bpy.context.view_layer.depsgraph,
v.co + cam_pos * 0.00000001, dir)
if not hit:
v.select_set(True)
selected_verts += 1
else:
v.select_set(False)
if selected_verts > 0:
visible_objs.append(obj.name)
print("all_verts: ", all_verts)
print("selected_verts: ", selected_verts)
bpy.ops.object.mode_set(mode='OBJECT')
print("Done visible objs are: ", visible_objs)
print("-----------------end---------------------")
The main issue that keeps your code from working correctly is that you are using the vertex coordinates in local space, but scene.raycast works off of world space coordinates. v.co is the vertex's position in local space, not world space. If you scale, rotate, and grab the object, that transformation is stored separately as world space transformation. You would need to convert the vertex's local coordinate to world space. To apply the world matrix transformation, write the coordinate as:
transformed_vertex_coordinate = obj.matrix_world @ v.co
dir = cam_pos - transformed_vertex_coordinate
hit, _, _, _, ob, _ = bpy.context.scene.ray_cast(bpy.context.view_layer.depsgraph, transformed_vertex_coordinate + cam_pos * 0.00000001, dir)
I already have code based on my add-ons nView and nView Live that I think you will find more useful and, in my opinion, more accurate. For any use case, you have to decide what determines visibility and accuracy. Vertex visibility is useful but there are cases where that may bring false negatives (say, if the vertices themselves are occluded or outside the camera frame, but their face is not). So I usually just do a low-resolution raycast from the camera. Combined with scene.ray_cast, users can check visibility independent of mesh resolution and customize their level of precision. Want more accuracy? Make the resolution higher. But you can get away with fewer rays, 25% or even 10% of screen resolution. Plus, it is much faster.
Enjoy:
import numpy as np
from mathutils import Vector
import bpy
def occlusion_test(scene, depsgraph, camera, resolution_x, resolution_y):
# get vectors which define view frustum of camera
top_right, _, bottom_left, top_left = camera.data.view_frame(scene=scene)
camera_quaternion = camera.matrix_world.to_quaternion()
camera_translation = camera.matrix_world.translation
# get iteration range for both the x and y axes, sampled based on the resolution
x_range = np.linspace(top_left[0], top_right[0], resolution_x)
y_range = np.linspace(top_left[1], bottom_left[1], resolution_y)
z_dir = top_left[2]
hit_data = set()
# iterate over all X/Y coordinates
for x in x_range:
for y in y_range:
# get current pixel vector from camera center to pixel
pixel_vector = Vector((x, y, z_dir))
# rotate that vector according to camera rotation
pixel_vector.rotate(camera_quaternion)
pixel_vector.normalized()
is_hit, _, _, _, hit_obj, _ = scene.ray_cast(depsgraph, camera_translation, pixel_vector)
if is_hit:
hit_data.add(hit_obj.name)
return hit_data
context = bpy.context
# sampling resolution of raytracing from the camera
# usually scene objects are not pixel-sized, so you can get away with fewer pixels
res_ratio = 0.25
res_x = int(context.scene.render.resolution_x * res_ratio)
res_y = int(context.scene.render.resolution_y * res_ratio)
visible_objs = occlusion_test(context.scene, context.evaluated_depsgraph_get(), context.scene.objects['Camera'],
res_x, res_y)
print('Visible objects:', visible_objs)
# if you want the objects NOT seen, ie occluded objects
invisible_objs = {o.name for o in context.scene.objects if o.type == 'MESH' and o.name not in visible_objs}
print('Invisible objects:', invisible_objs)
```
