Spaces:
Running
on
L40S
Running
on
L40S
update
Browse files- app.py +14 -14
- infer_api.py +68 -73
- refine/mesh_refine.py +168 -14
- slrm/models/lrm_mesh.py +2 -2
app.py
CHANGED
@@ -10,20 +10,20 @@ import os
|
|
10 |
import shlex
|
11 |
import subprocess
|
12 |
|
13 |
-
os.makedirs("./ckpt", exist_ok=True)
|
14 |
-
# download ViT-H SAM model into ./ckpt
|
15 |
-
subprocess.call(["wget", "-q", "https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth", "-O", "./ckpt/sam_vit_h_4b8939.pth"])
|
16 |
-
|
17 |
-
subprocess.run(
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
)
|
22 |
-
subprocess.run(
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
)
|
27 |
|
28 |
from infer_api import InferAPI
|
29 |
|
|
|
10 |
import shlex
|
11 |
import subprocess
|
12 |
|
13 |
+
# os.makedirs("./ckpt", exist_ok=True)
|
14 |
+
# # download ViT-H SAM model into ./ckpt
|
15 |
+
# subprocess.call(["wget", "-q", "https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth", "-O", "./ckpt/sam_vit_h_4b8939.pth"])
|
16 |
+
|
17 |
+
# subprocess.run(
|
18 |
+
# shlex.split(
|
19 |
+
# "pip install pip==24.0"
|
20 |
+
# )
|
21 |
+
# )
|
22 |
+
# subprocess.run(
|
23 |
+
# shlex.split(
|
24 |
+
# "pip install package/nvdiffrast-0.3.1.torch-cp310-cp310-linux_x86_64.whl --force-reinstall --no-deps"
|
25 |
+
# )
|
26 |
+
# )
|
27 |
|
28 |
from infer_api import InferAPI
|
29 |
|
infer_api.py
CHANGED
@@ -12,6 +12,7 @@ from omegaconf import OmegaConf
|
|
12 |
import numpy as np
|
13 |
|
14 |
import torch
|
|
|
15 |
|
16 |
from diffusers import AutoencoderKL, DDIMScheduler
|
17 |
from diffusers.utils import check_min_version
|
@@ -72,7 +73,7 @@ from slrm.utils.camera_util import (
|
|
72 |
FOV_to_intrinsics,
|
73 |
get_circular_camera_poses,
|
74 |
)
|
75 |
-
from slrm.utils.mesh_util import save_obj, save_glb
|
76 |
from slrm.utils.infer_util import images_to_video
|
77 |
|
78 |
import cv2
|
@@ -477,7 +478,7 @@ def calc_horizontal_offset2(target_mask, source_img):
|
|
477 |
|
478 |
|
479 |
@spaces.GPU
|
480 |
-
def get_distract_mask(
|
481 |
distract_area = np.abs(color_0 - color_1).sum(axis=-1) > thres
|
482 |
if normal_0 is not None and normal_1 is not None:
|
483 |
distract_area |= np.abs(normal_0 - normal_1).sum(axis=-1) > thres
|
@@ -503,43 +504,7 @@ def get_distract_mask(generator, color_0, color_1, normal_0=None, normal_1=None,
|
|
503 |
max_x, max_y = bbox.max(axis=0)
|
504 |
distract_bbox[min_x:max_x, min_y:max_y] = 1
|
505 |
|
506 |
-
|
507 |
-
labels = np.ones(len(points), dtype=np.int32)
|
508 |
-
|
509 |
-
masks = generator.generate((color_1 * 255).astype(np.uint8))
|
510 |
-
|
511 |
-
outside_area = np.abs(color_0 - color_1).sum(axis=-1) < outside_thres
|
512 |
-
|
513 |
-
final_mask = np.zeros_like(distract_mask)
|
514 |
-
for iii, mask in enumerate(masks):
|
515 |
-
mask['segmentation'] = cv2.resize(mask['segmentation'].astype(np.float32), (1024, 1024)) > 0.5
|
516 |
-
intersection = np.logical_and(mask['segmentation'], distract_mask).sum()
|
517 |
-
total = mask['segmentation'].sum()
|
518 |
-
iou = intersection / total
|
519 |
-
outside_intersection = np.logical_and(mask['segmentation'], outside_area).sum()
|
520 |
-
outside_total = mask['segmentation'].sum()
|
521 |
-
outside_iou = outside_intersection / outside_total
|
522 |
-
if iou > ratio and outside_iou < outside_ratio:
|
523 |
-
final_mask |= mask['segmentation']
|
524 |
-
|
525 |
-
# calculate coverage
|
526 |
-
intersection = np.logical_and(final_mask, distract_mask).sum()
|
527 |
-
total = distract_mask.sum()
|
528 |
-
coverage = intersection / total
|
529 |
-
|
530 |
-
if coverage < 0.8:
|
531 |
-
# use original distract mask
|
532 |
-
final_mask = (distract_mask.copy() * 255).astype(np.uint8)
|
533 |
-
final_mask = cv2.dilate(final_mask, np.ones((3, 3), np.uint8), iterations=3)
|
534 |
-
labeled_array_dilate, num_features_dilate = scipy.ndimage.label(final_mask)
|
535 |
-
for i in range(num_features_dilate + 1):
|
536 |
-
if np.sum(labeled_array_dilate == i) < 200:
|
537 |
-
final_mask[labeled_array_dilate == i] = 255
|
538 |
-
|
539 |
-
final_mask = cv2.erode(final_mask, np.ones((3, 3), np.uint8), iterations=3)
|
540 |
-
final_mask = final_mask > 127
|
541 |
-
|
542 |
-
return distract_mask, distract_bbox, random_sampled_points, final_mask
|
543 |
|
544 |
|
545 |
# infer_refine_sam = sam_model_registry["vit_h"](checkpoint="./ckpt/sam_vit_h_4b8939.pth").cuda()
|
@@ -563,6 +528,7 @@ def infer_refine(meshes, imgs):
|
|
563 |
distract_mask = None
|
564 |
|
565 |
results = []
|
|
|
566 |
|
567 |
for name_idx, level in zip([2, 0, 1], [2, 1, 0]):
|
568 |
mesh = trimesh.load(meshes[name_idx])
|
@@ -607,11 +573,11 @@ def infer_refine(meshes, imgs):
|
|
607 |
colors.append(color)
|
608 |
normals.append(normal)
|
609 |
|
610 |
-
|
611 |
-
|
612 |
-
|
613 |
-
|
614 |
-
|
615 |
|
616 |
if last_colors is None:
|
617 |
from copy import deepcopy
|
@@ -625,15 +591,15 @@ def infer_refine(meshes, imgs):
|
|
625 |
_, idx_anchor = kdtree_anchor.query(mesh_v, k=1)
|
626 |
_, idx_mesh_v = kdtree_mesh_v.query(mesh_v, k=25)
|
627 |
idx_anchor = idx_anchor.squeeze()
|
628 |
-
neighbors = torch.tensor(mesh_v)[idx_mesh_v] # V, 25, 3
|
629 |
# calculate the distances neighbors [V, 25, 3]; mesh_v [V, 3] -> [V, 25]
|
630 |
-
neighbor_dists = torch.norm(neighbors - torch.tensor(mesh_v)[:, None], dim=-1)
|
631 |
neighbor_dists[neighbor_dists > 0.06] = 114514.
|
632 |
neighbor_weights = torch.exp(-neighbor_dists * 1.)
|
633 |
neighbor_weights = neighbor_weights / neighbor_weights.sum(dim=1, keepdim=True)
|
634 |
anchors = fixed_v[idx_anchor] # V, 3
|
635 |
anchor_normals = calc_vertex_normals(fixed_v, fixed_f)[idx_anchor] # V, 3
|
636 |
-
dis_anchor = torch.clamp(((anchors - torch.tensor(mesh_v)) * anchor_normals).sum(-1), min=0) + 0.01
|
637 |
vec_anchor = dis_anchor[:, None] * anchor_normals # V, 3
|
638 |
vec_anchor = vec_anchor[idx_mesh_v] # V, 25, 3
|
639 |
weighted_vec_anchor = (vec_anchor * neighbor_weights[:, :, None]).sum(1) # V, 3
|
@@ -647,7 +613,7 @@ def infer_refine(meshes, imgs):
|
|
647 |
# my mesh flow weight by nearest vertexs
|
648 |
try:
|
649 |
if fixed_v is not None and fixed_f is not None and level != 0:
|
650 |
-
new_mesh_v = new_mesh.
|
651 |
|
652 |
fixed_v_cpu = fixed_v.cpu().numpy()
|
653 |
kdtree_anchor = KDTree(fixed_v_cpu)
|
@@ -655,48 +621,60 @@ def infer_refine(meshes, imgs):
|
|
655 |
_, idx_anchor = kdtree_anchor.query(new_mesh_v, k=1)
|
656 |
_, idx_mesh_v = kdtree_mesh_v.query(new_mesh_v, k=25)
|
657 |
idx_anchor = idx_anchor.squeeze()
|
658 |
-
neighbors = torch.tensor(new_mesh_v)[idx_mesh_v] # V, 25, 3
|
659 |
# calculate the distances neighbors [V, 25, 3]; new_mesh_v [V, 3] -> [V, 25]
|
660 |
-
neighbor_dists = torch.norm(neighbors - torch.tensor(new_mesh_v)[:, None], dim=-1)
|
661 |
neighbor_dists[neighbor_dists > 0.06] = 114514.
|
662 |
neighbor_weights = torch.exp(-neighbor_dists * 1.)
|
663 |
neighbor_weights = neighbor_weights / neighbor_weights.sum(dim=1, keepdim=True)
|
664 |
anchors = fixed_v[idx_anchor] # V, 3
|
665 |
anchor_normals = calc_vertex_normals(fixed_v, fixed_f)[idx_anchor] # V, 3
|
666 |
-
dis_anchor = torch.clamp(((anchors - torch.tensor(new_mesh_v)) * anchor_normals).sum(-1), min=0) + 0.01
|
667 |
vec_anchor = dis_anchor[:, None] * anchor_normals # V, 3
|
668 |
vec_anchor = vec_anchor[idx_mesh_v] # V, 25, 3
|
669 |
weighted_vec_anchor = (vec_anchor * neighbor_weights[:, :, None]).sum(1) # V, 3
|
670 |
new_mesh_v += weighted_vec_anchor.cpu().numpy()
|
671 |
|
672 |
# replace new_mesh verts with new_mesh_v
|
673 |
-
new_mesh =
|
674 |
|
675 |
except Exception as e:
|
676 |
pass
|
677 |
|
678 |
-
notsimp_v, notsimp_f, notsimp_t = new_mesh.verts_packed(), new_mesh.faces_packed(), new_mesh.textures.verts_features_packed()
|
679 |
-
|
680 |
if fixed_v is None:
|
681 |
fixed_v, fixed_f = simp_v, simp_f
|
682 |
-
complete_v, complete_f, complete_t = notsimp_v, notsimp_f, notsimp_t
|
683 |
else:
|
684 |
fixed_f = torch.cat([fixed_f, simp_f + fixed_v.shape[0]], dim=0)
|
685 |
fixed_v = torch.cat([fixed_v, simp_v], dim=0)
|
686 |
-
|
687 |
-
|
688 |
-
complete_v = torch.cat([complete_v, notsimp_v], dim=0)
|
689 |
-
complete_t = torch.cat([complete_t, notsimp_t], dim=0)
|
690 |
|
691 |
if level == 2:
|
692 |
-
new_mesh =
|
693 |
|
694 |
-
|
695 |
-
results.append(meshes[name_idx].replace('.obj', '_refined.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
696 |
|
697 |
# save whole mesh
|
698 |
-
|
699 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
700 |
|
701 |
return results
|
702 |
|
@@ -749,7 +727,7 @@ def infer_slrm_make3d(images):
|
|
749 |
return mesh_glb_fpaths
|
750 |
|
751 |
@spaces.GPU
|
752 |
-
def infer_slrm_make_mesh(mesh_fpath, planes, level=None):
|
753 |
mesh_basename = os.path.basename(mesh_fpath).split('.')[0]
|
754 |
mesh_dirname = os.path.dirname(mesh_fpath)
|
755 |
|
@@ -757,19 +735,36 @@ def infer_slrm_make_mesh(mesh_fpath, planes, level=None):
|
|
757 |
# get mesh
|
758 |
mesh_out = infer_slrm_model.extract_mesh(
|
759 |
planes,
|
760 |
-
use_texture_map=
|
761 |
levels=torch.tensor([level]).to(device),
|
762 |
**infer_slrm_infer_config,
|
763 |
)
|
764 |
|
765 |
-
|
766 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
767 |
|
768 |
-
|
769 |
-
|
770 |
-
|
771 |
-
|
772 |
-
|
773 |
|
774 |
return mesh_fpath
|
775 |
|
|
|
12 |
import numpy as np
|
13 |
|
14 |
import torch
|
15 |
+
from pygltflib import GLTF2, Material, PbrMetallicRoughness
|
16 |
|
17 |
from diffusers import AutoencoderKL, DDIMScheduler
|
18 |
from diffusers.utils import check_min_version
|
|
|
73 |
FOV_to_intrinsics,
|
74 |
get_circular_camera_poses,
|
75 |
)
|
76 |
+
from slrm.utils.mesh_util import save_obj, save_glb, save_obj_with_mtl
|
77 |
from slrm.utils.infer_util import images_to_video
|
78 |
|
79 |
import cv2
|
|
|
478 |
|
479 |
|
480 |
@spaces.GPU
|
481 |
+
def get_distract_mask(color_0, color_1, normal_0=None, normal_1=None, thres=0.25, ratio=0.50, outside_thres=0.10, outside_ratio=0.20):
|
482 |
distract_area = np.abs(color_0 - color_1).sum(axis=-1) > thres
|
483 |
if normal_0 is not None and normal_1 is not None:
|
484 |
distract_area |= np.abs(normal_0 - normal_1).sum(axis=-1) > thres
|
|
|
504 |
max_x, max_y = bbox.max(axis=0)
|
505 |
distract_bbox[min_x:max_x, min_y:max_y] = 1
|
506 |
|
507 |
+
return distract_mask, distract_bbox
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
508 |
|
509 |
|
510 |
# infer_refine_sam = sam_model_registry["vit_h"](checkpoint="./ckpt/sam_vit_h_4b8939.pth").cuda()
|
|
|
528 |
distract_mask = None
|
529 |
|
530 |
results = []
|
531 |
+
mesh_list = []
|
532 |
|
533 |
for name_idx, level in zip([2, 0, 1], [2, 1, 0]):
|
534 |
mesh = trimesh.load(meshes[name_idx])
|
|
|
573 |
colors.append(color)
|
574 |
normals.append(normal)
|
575 |
|
576 |
+
if last_front_color is not None and level == 0:
|
577 |
+
distract_mask, distract_bbox = get_distract_mask(last_front_color, np.array(colors[0]).astype(np.float32) / 255.0)
|
578 |
+
else:
|
579 |
+
distract_mask = None
|
580 |
+
distract_bbox = None
|
581 |
|
582 |
if last_colors is None:
|
583 |
from copy import deepcopy
|
|
|
591 |
_, idx_anchor = kdtree_anchor.query(mesh_v, k=1)
|
592 |
_, idx_mesh_v = kdtree_mesh_v.query(mesh_v, k=25)
|
593 |
idx_anchor = idx_anchor.squeeze()
|
594 |
+
neighbors = torch.tensor(mesh_v).cuda()[idx_mesh_v] # V, 25, 3
|
595 |
# calculate the distances neighbors [V, 25, 3]; mesh_v [V, 3] -> [V, 25]
|
596 |
+
neighbor_dists = torch.norm(neighbors - torch.tensor(mesh_v).cuda()[:, None], dim=-1)
|
597 |
neighbor_dists[neighbor_dists > 0.06] = 114514.
|
598 |
neighbor_weights = torch.exp(-neighbor_dists * 1.)
|
599 |
neighbor_weights = neighbor_weights / neighbor_weights.sum(dim=1, keepdim=True)
|
600 |
anchors = fixed_v[idx_anchor] # V, 3
|
601 |
anchor_normals = calc_vertex_normals(fixed_v, fixed_f)[idx_anchor] # V, 3
|
602 |
+
dis_anchor = torch.clamp(((anchors - torch.tensor(mesh_v).cuda()) * anchor_normals).sum(-1), min=0) + 0.01
|
603 |
vec_anchor = dis_anchor[:, None] * anchor_normals # V, 3
|
604 |
vec_anchor = vec_anchor[idx_mesh_v] # V, 25, 3
|
605 |
weighted_vec_anchor = (vec_anchor * neighbor_weights[:, :, None]).sum(1) # V, 3
|
|
|
613 |
# my mesh flow weight by nearest vertexs
|
614 |
try:
|
615 |
if fixed_v is not None and fixed_f is not None and level != 0:
|
616 |
+
new_mesh_v = new_mesh.vertices.copy()
|
617 |
|
618 |
fixed_v_cpu = fixed_v.cpu().numpy()
|
619 |
kdtree_anchor = KDTree(fixed_v_cpu)
|
|
|
621 |
_, idx_anchor = kdtree_anchor.query(new_mesh_v, k=1)
|
622 |
_, idx_mesh_v = kdtree_mesh_v.query(new_mesh_v, k=25)
|
623 |
idx_anchor = idx_anchor.squeeze()
|
624 |
+
neighbors = torch.tensor(new_mesh_v).cuda()[idx_mesh_v] # V, 25, 3
|
625 |
# calculate the distances neighbors [V, 25, 3]; new_mesh_v [V, 3] -> [V, 25]
|
626 |
+
neighbor_dists = torch.norm(neighbors - torch.tensor(new_mesh_v).cuda()[:, None], dim=-1)
|
627 |
neighbor_dists[neighbor_dists > 0.06] = 114514.
|
628 |
neighbor_weights = torch.exp(-neighbor_dists * 1.)
|
629 |
neighbor_weights = neighbor_weights / neighbor_weights.sum(dim=1, keepdim=True)
|
630 |
anchors = fixed_v[idx_anchor] # V, 3
|
631 |
anchor_normals = calc_vertex_normals(fixed_v, fixed_f)[idx_anchor] # V, 3
|
632 |
+
dis_anchor = torch.clamp(((anchors - torch.tensor(new_mesh_v).cuda()) * anchor_normals).sum(-1), min=0) + 0.01
|
633 |
vec_anchor = dis_anchor[:, None] * anchor_normals # V, 3
|
634 |
vec_anchor = vec_anchor[idx_mesh_v] # V, 25, 3
|
635 |
weighted_vec_anchor = (vec_anchor * neighbor_weights[:, :, None]).sum(1) # V, 3
|
636 |
new_mesh_v += weighted_vec_anchor.cpu().numpy()
|
637 |
|
638 |
# replace new_mesh verts with new_mesh_v
|
639 |
+
new_mesh.vertices = new_mesh_v
|
640 |
|
641 |
except Exception as e:
|
642 |
pass
|
643 |
|
|
|
|
|
644 |
if fixed_v is None:
|
645 |
fixed_v, fixed_f = simp_v, simp_f
|
|
|
646 |
else:
|
647 |
fixed_f = torch.cat([fixed_f, simp_f + fixed_v.shape[0]], dim=0)
|
648 |
fixed_v = torch.cat([fixed_v, simp_v], dim=0)
|
649 |
+
|
650 |
+
mesh_list.append(new_mesh)
|
|
|
|
|
651 |
|
652 |
if level == 2:
|
653 |
+
new_mesh = trimesh.Trimesh(simp_v.cpu().numpy(), simp_f.cpu().numpy(), process=False)
|
654 |
|
655 |
+
new_mesh.export(meshes[name_idx].replace('.obj', '_refined.glb'))
|
656 |
+
results.append(meshes[name_idx].replace('.obj', '_refined.glb'))
|
657 |
+
|
658 |
+
gltf = GLTF2().load(meshes[name_idx].replace('.obj', '_refined.glb'))
|
659 |
+
for material in gltf.materials:
|
660 |
+
if material.pbrMetallicRoughness:
|
661 |
+
material.pbrMetallicRoughness.baseColorFactor = [1.0, 1.0, 1.0, 100.0]
|
662 |
+
material.pbrMetallicRoughness.metallicFactor = 0.0
|
663 |
+
material.pbrMetallicRoughness.roughnessFactor = 1.0
|
664 |
+
gltf.save(meshes[name_idx].replace('.obj', '_refined.glb'))
|
665 |
|
666 |
# save whole mesh
|
667 |
+
scene = trimesh.Scene(mesh_list)
|
668 |
+
scene.export(meshes[name_idx].replace('.obj', '_refined_whole.glb'))
|
669 |
+
results.append(meshes[name_idx].replace('.obj', '_refined_whole.glb'))
|
670 |
+
|
671 |
+
gltf = GLTF2().load(meshes[name_idx].replace('.obj', '_refined_whole.glb'))
|
672 |
+
for material in gltf.materials:
|
673 |
+
if material.pbrMetallicRoughness:
|
674 |
+
material.pbrMetallicRoughness.baseColorFactor = [1.0, 1.0, 1.0, 100.0]
|
675 |
+
material.pbrMetallicRoughness.metallicFactor = 0.0
|
676 |
+
material.pbrMetallicRoughness.roughnessFactor = 1.0
|
677 |
+
gltf.save(meshes[name_idx].replace('.obj', '_refined_whole.glb'))
|
678 |
|
679 |
return results
|
680 |
|
|
|
727 |
return mesh_glb_fpaths
|
728 |
|
729 |
@spaces.GPU
|
730 |
+
def infer_slrm_make_mesh(mesh_fpath, planes, level=None, use_texture_map=False):
|
731 |
mesh_basename = os.path.basename(mesh_fpath).split('.')[0]
|
732 |
mesh_dirname = os.path.dirname(mesh_fpath)
|
733 |
|
|
|
735 |
# get mesh
|
736 |
mesh_out = infer_slrm_model.extract_mesh(
|
737 |
planes,
|
738 |
+
use_texture_map=use_texture_map,
|
739 |
levels=torch.tensor([level]).to(device),
|
740 |
**infer_slrm_infer_config,
|
741 |
)
|
742 |
|
743 |
+
if use_texture_map:
|
744 |
+
vertices, faces, uvs, mesh_tex_idx, tex_map = mesh_out
|
745 |
+
vertices = vertices[:, [1, 2, 0]]
|
746 |
+
tex_map = tex_map.permute(1, 2, 0).data.cpu().numpy()
|
747 |
+
|
748 |
+
if level == 2:
|
749 |
+
# fill all vertex_colors with 127
|
750 |
+
tex_map = np.ones_like(tex_map) * 127
|
751 |
+
save_obj_with_mtl(
|
752 |
+
vertices.data.cpu().numpy(),
|
753 |
+
uvs.data.cpu().numpy(),
|
754 |
+
faces.data.cpu().numpy(),
|
755 |
+
mesh_tex_idx.data.cpu().numpy(),
|
756 |
+
tex_map,
|
757 |
+
mesh_fpath
|
758 |
+
)
|
759 |
+
else:
|
760 |
+
vertices, faces, vertex_colors = mesh_out
|
761 |
+
vertices = vertices[:, [1, 2, 0]]
|
762 |
|
763 |
+
if level == 2:
|
764 |
+
# fill all vertex_colors with 127
|
765 |
+
vertex_colors = np.ones_like(vertex_colors) * 127
|
766 |
+
|
767 |
+
save_obj(vertices, faces, vertex_colors, mesh_fpath)
|
768 |
|
769 |
return mesh_fpath
|
770 |
|
refine/mesh_refine.py
CHANGED
@@ -13,6 +13,104 @@ from refine.render import NormalsRenderer, calc_vertex_normals
|
|
13 |
|
14 |
import pytorch3d
|
15 |
from pytorch3d.structures import Meshes
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
|
17 |
def remove_color(arr):
|
18 |
if arr.shape[-1] == 4:
|
@@ -301,11 +399,11 @@ def geo_refine_1(mesh_v, mesh_f, rgb_ls, normal_ls, expansion_weight=0.1, fixed_
|
|
301 |
return mesh_v, mesh_f
|
302 |
|
303 |
vertices, faces = reconstruct_stage1(rm_normals, steps=200, vertices=mesh_v, faces=mesh_f, fixed_v=fixed_v, fixed_f=fixed_f,
|
304 |
-
lr=stage1_lr, remesh_interval=stage1_remesh_interval, start_edge_len=0.
|
305 |
-
end_edge_len=0.
|
306 |
distract_mask=distract_mask, distract_bbox=distract_bbox)
|
307 |
|
308 |
-
vertices, faces = run_mesh_refine(vertices, faces, rm_normals, fixed_v=fixed_v, fixed_f=fixed_f, steps=100, start_edge_len=0.
|
309 |
decay=0.99, update_normal_interval=20, update_warmup=5, process_inputs=False, process_outputs=False, remesh_interval=1)
|
310 |
return vertices, faces
|
311 |
|
@@ -314,21 +412,77 @@ def geo_refine_2(vertices, faces, fixed_v=None):
|
|
314 |
simp_vertices, simp_faces = meshes.verts_packed(), meshes.faces_packed()
|
315 |
vertices, faces = simp_vertices.detach().cpu().numpy(), simp_faces.detach().cpu().numpy()
|
316 |
# vertices, faces = trimesh.remesh.subdivide(vertices, faces)
|
317 |
-
if fixed_v is not None:
|
318 |
-
vertices, faces = trimesh.remesh.subdivide(vertices, faces)
|
319 |
return vertices, faces
|
320 |
|
321 |
-
def geo_refine_3(
|
322 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
323 |
# concatenate fixed_v and fixed_f
|
324 |
if fixed_v is not None and fixed_f is not None:
|
325 |
-
|
326 |
-
|
327 |
# reconstruct meshes
|
328 |
-
meshes = Meshes(verts=[
|
329 |
new_meshes = multiview_color_projection(meshes, rgb_ls, resolution=1024, device="cuda", complete_unseen=True, confidence_threshold=0.2, cameras_list = get_cameras_list([180, 225, 270, 0, 90, 135], "cuda", focal=1/1.2), weights=[2.0, 0.5, 0.0, 1.0, 0.0, 0.5] if distract_mask is None else [2.0, 0.0, 0.5, 1.0, 0.5, 0.0], distract_mask=distract_mask)
|
330 |
-
|
331 |
if fixed_v is not None and fixed_f is not None:
|
332 |
-
|
333 |
-
|
334 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
|
14 |
import pytorch3d
|
15 |
from pytorch3d.structures import Meshes
|
16 |
+
import xatlas
|
17 |
+
import cv2
|
18 |
+
|
19 |
+
|
20 |
+
def mesh_uv_wrap(vertices, faces):
|
21 |
+
if len(faces) > 50000:
|
22 |
+
raise ValueError("The mesh has more than 50,000 faces, which is not supported.")
|
23 |
+
|
24 |
+
vmapping, indices, uvs = xatlas.parametrize(vertices, faces)
|
25 |
+
return vertices[vmapping], indices, uvs
|
26 |
+
|
27 |
+
|
28 |
+
def stride_from_shape(shape):
|
29 |
+
stride = [1]
|
30 |
+
for x in reversed(shape[1:]):
|
31 |
+
stride.append(stride[-1] * x)
|
32 |
+
return list(reversed(stride))
|
33 |
+
|
34 |
+
def scatter_add_nd_with_count(input, count, indices, values, weights=None):
|
35 |
+
# input: [..., C], D dimension + C channel
|
36 |
+
# count: [..., 1], D dimension
|
37 |
+
# indices: [N, D], long
|
38 |
+
# values: [N, C]
|
39 |
+
|
40 |
+
D = indices.shape[-1]
|
41 |
+
C = input.shape[-1]
|
42 |
+
size = input.shape[:-1]
|
43 |
+
stride = stride_from_shape(size)
|
44 |
+
|
45 |
+
assert len(size) == D
|
46 |
+
|
47 |
+
input = input.view(-1, C) # [HW, C]
|
48 |
+
count = count.view(-1, 1)
|
49 |
+
|
50 |
+
flatten_indices = (indices * torch.tensor(stride,
|
51 |
+
dtype=torch.long, device=indices.device)).sum(-1) # [N]
|
52 |
+
|
53 |
+
if weights is None:
|
54 |
+
weights = torch.ones_like(values[..., :1])
|
55 |
+
|
56 |
+
input.scatter_add_(0, flatten_indices.unsqueeze(1).repeat(1, C), values)
|
57 |
+
count.scatter_add_(0, flatten_indices.unsqueeze(1), weights)
|
58 |
+
|
59 |
+
return input.view(*size, C), count.view(*size, 1)
|
60 |
+
|
61 |
+
|
62 |
+
def linear_grid_put_2d(H, W, coords, values, return_count=False):
|
63 |
+
# coords: [N, 2], float in [0, 1]
|
64 |
+
# values: [N, C]
|
65 |
+
|
66 |
+
C = values.shape[-1]
|
67 |
+
|
68 |
+
indices = coords * torch.tensor(
|
69 |
+
[H - 1, W - 1], dtype=torch.float32, device=coords.device
|
70 |
+
)
|
71 |
+
indices_00 = indices.floor().long() # [N, 2]
|
72 |
+
indices_00[:, 0].clamp_(0, H - 2)
|
73 |
+
indices_00[:, 1].clamp_(0, W - 2)
|
74 |
+
indices_01 = indices_00 + torch.tensor(
|
75 |
+
[0, 1], dtype=torch.long, device=indices.device
|
76 |
+
)
|
77 |
+
indices_10 = indices_00 + torch.tensor(
|
78 |
+
[1, 0], dtype=torch.long, device=indices.device
|
79 |
+
)
|
80 |
+
indices_11 = indices_00 + torch.tensor(
|
81 |
+
[1, 1], dtype=torch.long, device=indices.device
|
82 |
+
)
|
83 |
+
|
84 |
+
h = indices[..., 0] - indices_00[..., 0].float()
|
85 |
+
w = indices[..., 1] - indices_00[..., 1].float()
|
86 |
+
w_00 = (1 - h) * (1 - w)
|
87 |
+
w_01 = (1 - h) * w
|
88 |
+
w_10 = h * (1 - w)
|
89 |
+
w_11 = h * w
|
90 |
+
|
91 |
+
result = torch.zeros(H, W, C, device=values.device,
|
92 |
+
dtype=values.dtype) # [H, W, C]
|
93 |
+
count = torch.zeros(H, W, 1, device=values.device,
|
94 |
+
dtype=values.dtype) # [H, W, 1]
|
95 |
+
weights = torch.ones_like(values[..., :1]) # [N, 1]
|
96 |
+
|
97 |
+
result, count = scatter_add_nd_with_count(
|
98 |
+
result, count, indices_00, values * w_00.unsqueeze(1), weights * w_00.unsqueeze(1))
|
99 |
+
result, count = scatter_add_nd_with_count(
|
100 |
+
result, count, indices_01, values * w_01.unsqueeze(1), weights * w_01.unsqueeze(1))
|
101 |
+
result, count = scatter_add_nd_with_count(
|
102 |
+
result, count, indices_10, values * w_10.unsqueeze(1), weights * w_10.unsqueeze(1))
|
103 |
+
result, count = scatter_add_nd_with_count(
|
104 |
+
result, count, indices_11, values * w_11.unsqueeze(1), weights * w_11.unsqueeze(1))
|
105 |
+
|
106 |
+
if return_count:
|
107 |
+
return result, count
|
108 |
+
|
109 |
+
mask = (count.squeeze(-1) > 0)
|
110 |
+
result[mask] = result[mask] / count[mask].repeat(1, C)
|
111 |
+
|
112 |
+
return result, count.squeeze(-1) == 0
|
113 |
+
|
114 |
|
115 |
def remove_color(arr):
|
116 |
if arr.shape[-1] == 4:
|
|
|
399 |
return mesh_v, mesh_f
|
400 |
|
401 |
vertices, faces = reconstruct_stage1(rm_normals, steps=200, vertices=mesh_v, faces=mesh_f, fixed_v=fixed_v, fixed_f=fixed_f,
|
402 |
+
lr=stage1_lr, remesh_interval=stage1_remesh_interval, start_edge_len=0.04,
|
403 |
+
end_edge_len=0.02, gain=0.05, loss_expansion_weight=expansion_weight,
|
404 |
distract_mask=distract_mask, distract_bbox=distract_bbox)
|
405 |
|
406 |
+
vertices, faces = run_mesh_refine(vertices, faces, rm_normals, fixed_v=fixed_v, fixed_f=fixed_f, steps=100, start_edge_len=0.02, end_edge_len=0.001,
|
407 |
decay=0.99, update_normal_interval=20, update_warmup=5, process_inputs=False, process_outputs=False, remesh_interval=1)
|
408 |
return vertices, faces
|
409 |
|
|
|
412 |
simp_vertices, simp_faces = meshes.verts_packed(), meshes.faces_packed()
|
413 |
vertices, faces = simp_vertices.detach().cpu().numpy(), simp_faces.detach().cpu().numpy()
|
414 |
# vertices, faces = trimesh.remesh.subdivide(vertices, faces)
|
|
|
|
|
415 |
return vertices, faces
|
416 |
|
417 |
+
def geo_refine_3(vertices_, faces_, rgb_ls, fixed_v=None, fixed_f=None, distract_mask=None):
|
418 |
+
# vertices, faces, uvs = mesh_uv_wrap(vertices_, faces_)
|
419 |
+
vmapping, indices, uvs = xatlas.parametrize(vertices_, faces_)
|
420 |
+
vertices, faces = vertices_[vmapping], indices
|
421 |
+
|
422 |
+
def subdivide(vertices, faces, uvs):
|
423 |
+
vertices, faces = trimesh.remesh.subdivide(
|
424 |
+
vertices=np.hstack((vertices, uvs.copy())),
|
425 |
+
faces=faces
|
426 |
+
)
|
427 |
+
return vertices[:, :3], faces, vertices[:, 3:]
|
428 |
+
|
429 |
+
if fixed_v is not None:
|
430 |
+
dense_atlas_vertices, dense_atlas_faces, dense_atlas_uvs = subdivide(vertices, faces, uvs)
|
431 |
+
dense_atlas_vertices, dense_atlas_faces, dense_atlas_uvs = subdivide(dense_atlas_vertices, dense_atlas_faces, dense_atlas_uvs)
|
432 |
+
# dense_atlas_vertices, dense_atlas_faces, dense_atlas_uvs = subdivide(dense_atlas_vertices, dense_atlas_faces, dense_atlas_uvs)
|
433 |
+
dense_vertices, dense_faces = trimesh.remesh.subdivide(vertices_, faces_)
|
434 |
+
dense_vertices, dense_faces = trimesh.remesh.subdivide(dense_vertices, dense_faces)
|
435 |
+
# dense_vertices, dense_faces = trimesh.remesh.subdivide(dense_vertices, dense_faces)
|
436 |
+
else:
|
437 |
+
dense_atlas_vertices, dense_atlas_faces, dense_atlas_uvs = subdivide(vertices, faces, uvs)
|
438 |
+
dense_atlas_vertices, dense_atlas_faces, dense_atlas_uvs = subdivide(dense_atlas_vertices, dense_atlas_faces, dense_atlas_uvs)
|
439 |
+
dense_vertices, dense_faces = trimesh.remesh.subdivide(vertices_, faces_)
|
440 |
+
dense_vertices, dense_faces = trimesh.remesh.subdivide(dense_vertices, dense_faces)
|
441 |
+
|
442 |
+
origin_len_v, origin_len_f = len(dense_vertices), len(dense_faces)
|
443 |
+
|
444 |
# concatenate fixed_v and fixed_f
|
445 |
if fixed_v is not None and fixed_f is not None:
|
446 |
+
dense_vertices, dense_faces = np.concatenate([dense_vertices, fixed_v.detach().cpu().numpy()], axis=0), np.concatenate([dense_faces, fixed_f.detach().cpu().numpy() + len(vertices)], axis=0)
|
447 |
+
dense_vertices, dense_faces = torch.from_numpy(dense_vertices).cuda(), torch.from_numpy(dense_faces.astype('int32')).cuda()
|
448 |
# reconstruct meshes
|
449 |
+
meshes = Meshes(verts=[dense_vertices], faces=[dense_faces], textures=pytorch3d.renderer.mesh.textures.TexturesVertex([torch.zeros_like(dense_vertices).float()]))
|
450 |
new_meshes = multiview_color_projection(meshes, rgb_ls, resolution=1024, device="cuda", complete_unseen=True, confidence_threshold=0.2, cameras_list = get_cameras_list([180, 225, 270, 0, 90, 135], "cuda", focal=1/1.2), weights=[2.0, 0.5, 0.0, 1.0, 0.0, 0.5] if distract_mask is None else [2.0, 0.0, 0.5, 1.0, 0.5, 0.0], distract_mask=distract_mask)
|
451 |
+
|
452 |
if fixed_v is not None and fixed_f is not None:
|
453 |
+
dense_vertices = dense_vertices[:origin_len_v]
|
454 |
+
dense_faces = dense_faces[:origin_len_f]
|
455 |
+
textures = new_meshes.textures.verts_features_packed()[:origin_len_v]
|
456 |
+
else:
|
457 |
+
textures = new_meshes.textures.verts_features_packed()
|
458 |
+
|
459 |
+
# distances = torch.cdist(torch.tensor(dense_atlas_vertices).cuda(), torch.tensor(dense_vertices).cuda())
|
460 |
+
# nearest_indices = torch.argmin(distances, dim=1)
|
461 |
+
# atlas_textures = textures[nearest_indices]
|
462 |
+
|
463 |
+
chunk_size = 500
|
464 |
+
atlas_textures_chunks = []
|
465 |
+
for i in range(0, len(dense_atlas_vertices), chunk_size):
|
466 |
+
chunk = dense_atlas_vertices[i:i+chunk_size]
|
467 |
+
distances = torch.cdist(torch.tensor(chunk).cuda(), torch.tensor(dense_vertices).cuda())
|
468 |
+
nearest_indices = torch.argmin(distances, dim=1)
|
469 |
+
atlas_textures_chunks.append(textures[nearest_indices])
|
470 |
+
atlas_textures = torch.cat(atlas_textures_chunks, dim=0)
|
471 |
+
|
472 |
+
dense_atlas_uvs = torch.tensor(dense_atlas_uvs, dtype=torch.float32).cuda()
|
473 |
+
tex_img, mask = linear_grid_put_2d(1024, 1024, dense_atlas_uvs, atlas_textures)
|
474 |
+
tex_img, mask = tex_img.cpu().numpy(), mask.cpu().numpy()
|
475 |
+
tex_img = cv2.inpaint((tex_img * 255).astype(np.uint8), (mask*255).astype('uint8'), 3, cv2.INPAINT_NS)
|
476 |
+
tex_img = Image.fromarray(np.transpose(tex_img,(1,0,2))[::-1])
|
477 |
+
|
478 |
+
mesh = trimesh.Trimesh(vertices, faces, process=False)
|
479 |
+
# material = trimesh.visual.texture.SimpleMaterial(image=tex_img, diffuse=(255, 255, 255))
|
480 |
+
material = trimesh.visual.material.PBRMaterial(
|
481 |
+
roughnessFactor=1.0,
|
482 |
+
baseColorTexture=tex_img,
|
483 |
+
baseColorFactor=np.array([255, 255, 255, 255], dtype=np.uint8)
|
484 |
+
)
|
485 |
+
texture_visuals = trimesh.visual.TextureVisuals(uv=uvs, image=tex_img, material=material)
|
486 |
+
mesh.visual = texture_visuals
|
487 |
+
|
488 |
+
return mesh, torch.tensor(vertices).cuda(), torch.tensor(faces.astype('int64')).cuda()
|
slrm/models/lrm_mesh.py
CHANGED
@@ -116,13 +116,13 @@ class MeshSLRM(nn.Module):
|
|
116 |
camera = OrthogonalCamera(device=device)
|
117 |
|
118 |
with torch.cuda.amp.autocast(enabled=False):
|
119 |
-
|
120 |
self.geometry = FlexiCubesGeometry(
|
121 |
grid_res_xy=self.grid_res_xy,
|
122 |
grid_res_z=self.grid_res_z,
|
123 |
scale_xy=self.grid_scale_xy,
|
124 |
scale_z=self.grid_scale_z,
|
125 |
-
renderer=
|
126 |
render_type='neural_render',
|
127 |
device=device,
|
128 |
)
|
|
|
116 |
camera = OrthogonalCamera(device=device)
|
117 |
|
118 |
with torch.cuda.amp.autocast(enabled=False):
|
119 |
+
renderer = NeuralRender(device, camera_model=camera)
|
120 |
self.geometry = FlexiCubesGeometry(
|
121 |
grid_res_xy=self.grid_res_xy,
|
122 |
grid_res_z=self.grid_res_z,
|
123 |
scale_xy=self.grid_scale_xy,
|
124 |
scale_z=self.grid_scale_z,
|
125 |
+
renderer=renderer,
|
126 |
render_type='neural_render',
|
127 |
device=device,
|
128 |
)
|