Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Make object acquisition demo callable by webapp #81

Merged
merged 1 commit into from
Jul 25, 2024
Merged

Make object acquisition demo callable by webapp #81

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
307 changes: 161 additions & 146 deletions scripts/acquire_object_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,154 +8,169 @@

b3d.rr_init("acquire_object_model")

# python scripts/acquire_object_model.py assets/shared_data_bucket/input_data/lysol_static.r3d

parser = argparse.ArgumentParser("acquire_object_mode")
parser.add_argument("input", help="r3d file", type=str)
args = parser.parse_args()

filename = args.input
data = b3d.io.load_r3d(filename)


_, _, fx, fy, cx, cy, near, far = data["camera_intrinsics_depth"]
image_height, image_width = data["depth"].shape[1:3]
num_scenes = data["depth"].shape[0]

indices = jnp.arange(0, num_scenes, 10)

camera_poses_full = data["camera_pose"]
camera_poses = camera_poses_full[indices]

xyz = b3d.xyz_from_depth_vectorized(data["depth"][indices], fx, fy, cx, cy)
xyz_world_frame = camera_poses[:, None, None].apply(xyz)

# for i in range(len(xyz_world_frame)):
# b3d.rr_set_time(i)
# b3d.utils.rr_log_cloud("xyz", xyz_world_frame[i])

# Resize rgbs to be same size as depth.
rgbs = data["rgb"]
rgbs_resized = jnp.clip(
jax.vmap(jax.image.resize, in_axes=(0, None, None))(
rgbs[indices] / 255.0,
(image_height, image_width, 3),
"linear",
),
0.0,
1.0,
)


masks = [b3d.carvekit_get_foreground_mask(r) for r in rgbs_resized]
masks_concat = jnp.stack(masks, axis=0)

grid_center = jnp.median(camera_poses[0].apply(xyz[0][masks[0]]), axis=0)
W = 0.3
D = 100
grid = jnp.stack(
jnp.meshgrid(
jnp.linspace(grid_center[0] - W / 2, grid_center[0] + W / 2, D),
jnp.linspace(grid_center[1] - W / 2, grid_center[1] + W / 2, D),
jnp.linspace(grid_center[2] - W / 2, grid_center[2] + W / 2, D),
),
axis=-1,
).reshape(-1, 3)

occ_free_occl_, colors_per_voxel_ = (
b3d.voxel_occupied_occluded_free_parallel_camera_depth(
camera_poses,
rgbs_resized,
xyz[..., 2] * masks_concat + (1.0 - masks_concat) * 5.0,
grid,
fx,
fy,
cx,
cy,
6.0,
0.005,

# ssh sam-b3d-l4.us-west1-a.probcomp-caliban -L 5000:localhost:5000


def acquire(input_path, output_path=None):
if output_path is None:
output_path = input_path + ".graphics_edits.mp4"

data = b3d.io.load_r3d(input_path)

_, _, fx, fy, cx, cy, near, far = data["camera_intrinsics_depth"]
image_height, image_width = data["depth"].shape[1:3]
num_scenes = data["depth"].shape[0]

indices = jnp.arange(0, num_scenes, 10)

camera_poses_full = data["camera_pose"]
camera_poses = camera_poses_full[indices]

xyz = b3d.xyz_from_depth_vectorized(data["depth"][indices], fx, fy, cx, cy)
xyz_world_frame = camera_poses[:, None, None].apply(xyz)

# for i in range(len(xyz_world_frame)):
# b3d.rr_set_time(i)
# b3d.utils.rr_log_cloud("xyz", xyz_world_frame[i])

# Resize rgbs to be same size as depth.
rgbs = data["rgb"]
rgbs_resized = jnp.clip(
jax.vmap(jax.image.resize, in_axes=(0, None, None))(
rgbs[indices] / 255.0,
(image_height, image_width, 3),
"linear",
),
0.0,
1.0,
)

masks = [b3d.carvekit_get_foreground_mask(r) for r in rgbs_resized]
masks_concat = jnp.stack(masks, axis=0)

grid_center = jnp.median(camera_poses[0].apply(xyz[0][masks[0]]), axis=0)
W = 0.3
D = 100
grid = jnp.stack(
jnp.meshgrid(
jnp.linspace(grid_center[0] - W / 2, grid_center[0] + W / 2, D),
jnp.linspace(grid_center[1] - W / 2, grid_center[1] + W / 2, D),
jnp.linspace(grid_center[2] - W / 2, grid_center[2] + W / 2, D),
),
axis=-1,
).reshape(-1, 3)

occ_free_occl_, colors_per_voxel_ = (
b3d.voxel_occupied_occluded_free_parallel_camera_depth(
camera_poses,
rgbs_resized,
xyz[..., 2] * masks_concat + (1.0 - masks_concat) * 5.0,
grid,
fx,
fy,
cx,
cy,
6.0,
0.005,
)
)
i = len(occ_free_occl_)
occ_free_occl, colors_per_voxel = occ_free_occl_[:i], colors_per_voxel_[:i]
total_occ = (occ_free_occl == 1.0).sum(0)
total_free = (occ_free_occl == -1.0).sum(0)
ratio = total_occ / (total_occ + total_free) * ((total_occ + total_free) > 1)

grid_colors = colors_per_voxel.sum(0) / (total_occ[..., None])
model_mask = ratio > 0.2

resolution = 0.0015

grid_points = grid[model_mask]
colors = grid_colors[model_mask]

meshes = b3d.mesh.transform_mesh(
jax.vmap(b3d.mesh.Mesh.cube_mesh)(
jnp.ones((grid_points.shape[0], 3)) * resolution * 2.0, colors
),
b3d.Pose.from_translation(grid_points)[:, None],
)
_object_mesh = b3d.mesh.Mesh.squeeze_mesh(meshes)

object_pose = Pose.from_translation(jnp.median(_object_mesh.vertices, axis=0))
object_mesh = _object_mesh.transform(object_pose.inv())
object_mesh.rr_visualize("mesh")

mesh_filename = input_path + ".mesh.obj"
# Save the mesh
print(f"Saving obj file to {mesh_filename}")
object_mesh.save(mesh_filename)

renderer = b3d.RendererOriginal(
image_width, image_height, fx, fy, cx, cy, near, far
)
rgbds = renderer.render_rgbd_many(
(camera_poses[:, None].inv() @ object_pose).apply(object_mesh.vertices),
object_mesh.faces,
jnp.tile(object_mesh.vertex_attributes, (len(camera_poses), 1, 1)),
)

sub_indices = jnp.array([0, 5, len(camera_poses) - 15, len(camera_poses) - 5])
mask = rgbds[sub_indices, ..., 3] == 0.0

background_xyzs = xyz_world_frame[sub_indices][mask]
colors = rgbs_resized[sub_indices][mask, :]
distances_from_camera = xyz[sub_indices][..., 2][mask][..., None] / fx

# subset = jax.random.choice(jax.random.PRNGKey(0), jnp.arange(background_xyzs.shape[0]), shape=(background_xyzs.shape[0]//3,), replace=False)

# background_xyzs = background_xyzs[subset]
# colors = colors[subset]
# distances_from_camera = distances_from_camera[subset]

meshes = b3d.mesh.transform_mesh(
jax.vmap(b3d.mesh.Mesh.cube_mesh)(
jnp.ones((background_xyzs.shape[0], 3)) * distances_from_camera, colors
),
b3d.Pose.from_translation(background_xyzs)[:, None],
)
)
i = len(occ_free_occl_)
occ_free_occl, colors_per_voxel = occ_free_occl_[:i], colors_per_voxel_[:i]
total_occ = (occ_free_occl == 1.0).sum(0)
total_free = (occ_free_occl == -1.0).sum(0)
ratio = total_occ / (total_occ + total_free) * ((total_occ + total_free) > 1)

grid_colors = colors_per_voxel.sum(0) / (total_occ[..., None])
model_mask = ratio > 0.2

resolution = 0.0015

grid_points = grid[model_mask]
colors = grid_colors[model_mask]

meshes = b3d.mesh.transform_mesh(
jax.vmap(b3d.mesh.Mesh.cube_mesh)(
jnp.ones((grid_points.shape[0], 3)) * resolution * 2.0, colors
),
b3d.Pose.from_translation(grid_points)[:, None],
)
_object_mesh = b3d.mesh.Mesh.squeeze_mesh(meshes)

object_pose = Pose.from_translation(jnp.median(_object_mesh.vertices, axis=0))
object_mesh = _object_mesh.transform(object_pose.inv())
object_mesh.rr_visualize("mesh")

mesh_filename = filename + ".mesh.obj"
# Save the mesh
print(f"Saving obj file to {mesh_filename}")
object_mesh.save(mesh_filename)

renderer = b3d.RendererOriginal(image_width, image_height, fx, fy, cx, cy, near, far)
rgbds = renderer.render_rgbd_many(
(camera_poses[:, None].inv() @ object_pose).apply(object_mesh.vertices),
object_mesh.faces,
jnp.tile(object_mesh.vertex_attributes, (len(camera_poses), 1, 1)),
)

sub_indices = jnp.array([0, 5, len(camera_poses) - 15, len(camera_poses) - 5])
mask = rgbds[sub_indices, ..., 3] == 0.0

background_xyzs = xyz_world_frame[sub_indices][mask]
colors = rgbs_resized[sub_indices][mask, :]
distances_from_camera = xyz[sub_indices][..., 2][mask][..., None] / fx

# subset = jax.random.choice(jax.random.PRNGKey(0), jnp.arange(background_xyzs.shape[0]), shape=(background_xyzs.shape[0]//3,), replace=False)

# background_xyzs = background_xyzs[subset]
# colors = colors[subset]
# distances_from_camera = distances_from_camera[subset]

meshes = b3d.mesh.transform_mesh(
jax.vmap(b3d.mesh.Mesh.cube_mesh)(
jnp.ones((background_xyzs.shape[0], 3)) * distances_from_camera, colors
),
b3d.Pose.from_translation(background_xyzs)[:, None],
)
background_mesh = b3d.mesh.Mesh.squeeze_mesh(meshes)
background_mesh.rr_visualize("background_mesh")


object_poses = [
object_pose,
Pose.identity(),
object_pose @ Pose.from_translation(jnp.array([-0.1, 0.0, 0.1])),
object_pose @ Pose.from_translation(jnp.array([-0.1, 0.0, -0.1])),
]

scene_mesh = b3d.mesh.transform_and_merge_meshes(
[object_mesh, background_mesh, object_mesh, object_mesh],
object_poses,
)

viz_images = []
for t in tqdm(range(len(camera_poses_full))):
b3d.utils.rr_set_time(t)
rgbd = renderer.render_rgbd_from_mesh(
scene_mesh.transform(camera_poses_full[t].inv())
background_mesh = b3d.mesh.Mesh.squeeze_mesh(meshes)
background_mesh.rr_visualize("background_mesh")

object_poses = [
object_pose,
Pose.identity(),
object_pose @ Pose.from_translation(jnp.array([-0.1, 0.0, 0.1])),
object_pose @ Pose.from_translation(jnp.array([-0.1, 0.0, -0.1])),
]

scene_mesh = b3d.mesh.transform_and_merge_meshes(
[object_mesh, background_mesh, object_mesh, object_mesh],
object_poses,
)
viz_images.append(b3d.viz_rgb(rgbd))

viz_images = []
for t in tqdm(range(len(camera_poses_full))):
b3d.utils.rr_set_time(t)
rgbd = renderer.render_rgbd_from_mesh(
scene_mesh.transform(camera_poses_full[t].inv())
)
viz_images.append(b3d.viz_rgb(rgbd))

b3d.make_video_from_pil_images(viz_images, output_path, fps=30.0)
print(f"Saved video to {output_path}")
return output_path


def main():
parser = argparse.ArgumentParser("acquire_object_mode")
parser.add_argument("input", help="r3d file", type=str)
args = parser.parse_args()
filename = args.input
return acquire(filename)


b3d.make_video_from_pil_images(viz_images, filename + ".graphics_edits.mp4", fps=30.0)
print(f"Saved video to {filename + '.graphics_edits.mp4'}")
if __name__ == "__main__":
main()
Loading