Repository navigation
Expand file tree
/
Copy patheval_shape_image_pair.py
More file actions
471 lines (380 loc) · 20.5 KB
/
Copy patheval_shape_image_pair.py
File metadata and controls
471 lines (380 loc) · 20.5 KB
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
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
import argparse
import json
import os
import random
import re
import cv2
import kaolin as kal
import numpy as np
import torch
from PIL import Image
from modules.utils import device, get_camera_from_view2, setcolor_mesh, setcolor_mesh_batched, show_mask, loadmesh
from modules.render import Renderer
from modules.sam import load_img_seg_model
from modules.best_segmentation_buddies import BestSegmentationBuddies
# Elevation angles searched when looking for a view that sees the target vertex.
ELEV_CANDIDATES = [-np.pi / 2, -np.pi / 3, -np.pi / 6, 0, np.pi / 6, np.pi / 3, np.pi / 2]
GREY = [2.0 / 3.0, 2.0 / 3.0, 2.0 / 3.0]
RED = [255.0, 0.0, 0.0]
class DepthControlNetGenerator:
def __init__(self, flux_model_path, device_str="cuda", seed=42, deterministic=True):
from diffusers import FluxControlPipeline
self.device = device_str
self.pipe = FluxControlPipeline.from_pretrained(
flux_model_path, torch_dtype=torch.bfloat16
).to(device_str)
# Deterministic: reuse `seed` every run, so the same view/prompt/depth
# map reproduces the same generated image. Random: draw a fresh seed
# each run.
if not deterministic:
seed = random.randint(0, 2**31 - 1)
print("[stage1] FLUX generation seed=%d (deterministic=%s)" % (seed, deterministic))
self.generator = torch.Generator(device=device_str).manual_seed(seed)
def generate(self, prompt, depth_estimation, num_inference_steps=30, guidance_scale=10.0):
"""Synthesize a natural image conditioned on `depth_estimation`."""
images = self.pipe(
prompt=prompt,
control_image=depth_estimation,
height=depth_estimation.size[1],
width=depth_estimation.size[0],
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
generator=self.generator,
output_type="pt",
).images
return images # [1, 3, H, W]
def to_pixel_coordinates(normalized_coords, width, height):
return torch.stack([(normalized_coords[..., 0] + 1) * 0.5 * width,
(normalized_coords[..., 1] + 1) * 0.5 * height], dim=-1)
def infer_shape_name(obj_path):
stem = os.path.splitext(os.path.basename(obj_path))[0]
prefix_match = re.match(r'^([a-zA-Z]+)', stem)
if prefix_match:
return prefix_match.group(1)
grandparent = os.path.basename(os.path.dirname(os.path.dirname(obj_path)))
return grandparent or stem
def find_visible_view(mesh, vertex, renderer, render_res, frontview_std, background):
vertex_rgb = torch.tensor(GREY, device=device).repeat(mesh.vertices.shape[0], 1)
vertex_rgb[vertex] = torch.tensor(RED, device=device)
setcolor_mesh(mesh, vertex_rgb)
elevs = list(ELEV_CANDIDATES)
azims = np.arange(0, 2 * np.pi, np.pi / 6).tolist()
random.shuffle(elevs)
random.shuffle(azims)
bg = torch.tensor(background).to(device)
for elev_c in elevs:
for azim_c in azims:
images, elev, azim, mask, verts_cam, verts_img, face_normals_z = renderer.render_views(
mesh, num_views=1, random_views=False,
center_elev=torch.tensor([elev_c]), center_azim=torch.tensor([azim_c]),
std=frontview_std, return_views=True, return_mask=True,
return_coordinates=True, lighting=False, background=bg)
verts_pix = to_pixel_coordinates(verts_img, render_res - 1, render_res - 1).round().long()
x = verts_pix[0][vertex][0].item()
y_raw = verts_pix[0][vertex][1].item()
row = render_res - 1 - y_raw # flip applied once, here
mask_bool = mask.squeeze(0).bool()
H, W = mask_bool.shape
if not (0 <= x < W and 0 <= row < H):
continue
# (a) neighborhood around the vertex must lie
# inside the object mask.
x0, x1 = max(0, x - 10), min(W - 1, x + 10)
r0a, r1a = max(0, row - 10), min(H - 1, row + 10)
if not mask_bool[r0a:r1a + 1, x0:x1 + 1].all():
continue
# (b) the vertex must be front-facing.
if torch.mean(face_normals_z[0][mesh.rows_for_each_vertex[vertex]]) < 0:
continue
# (c) among all vertices projecting to (x, y_raw), the target must
# be the closest to the camera.
overlaps = torch.where((verts_pix[0][:, 0] == x) & (verts_pix[0][:, 1] == y_raw))[0]
closest = overlaps[torch.argmax(verts_cam[0][overlaps][:, -1])]
if closest.item() != vertex:
continue
# (d) the marked (red) vertex must actually render red at image
# row = render_res - 1 - y_raw (the tensor's row convention).
px = images[0][:, row, x]
px_shift = px - px.min()
px_norm = px_shift / px_shift.max()
if not (px_norm == torch.tensor([1.0, 0.0, 0.0], device=device)).all():
continue
# (e) a small window around (row, x) must have at least some mask
# coverage.
r0, r1 = max(0, row - 2), min(H - 1, row + 2)
xx0, xx1 = max(0, x - 2), min(W - 1, x + 2)
if not mask_bool[r0:r1 + 1, xx0:xx1 + 1].any():
continue
return float(elev.item()), float(azim.item()), int(x), int(row)
return None
def render_clean_view(mesh, elev, azim, renderer, frontview_std, background):
vertex_rgb = torch.tensor(GREY, device=device).repeat(mesh.vertices.shape[0], 1)
setcolor_mesh(mesh, vertex_rgb)
images, _, _ = renderer.render_views(
mesh, num_views=1, random_views=False,
center_elev=torch.tensor([elev]), center_azim=torch.tensor([azim]),
std=frontview_std, return_views=True, lighting=True,
background=torch.tensor(background).to(device))
return images # [1, 3, render_res, render_res]
def render_depth(mesh, elev, azim, renderer):
faces = mesh.faces.to(device)
camera_transform = get_camera_from_view2(
torch.as_tensor(elev), torch.as_tensor(azim), r=renderer.radius).to(device)
face_vertices_camera, face_vertices_image, face_normals = kal.render.mesh.prepare_vertices(
mesh.vertices.to(device), faces, renderer.camera_projection, camera_transform=camera_transform)
depth_z = face_vertices_camera[:, :, :, -1]
depth_features = depth_z.unsqueeze(-1).repeat(1, 1, 1, 3)
depth_image, _, face_idx = kal.render.mesh.dibr_rasterization(
renderer.dim[1], renderer.dim[0], depth_z, face_vertices_image, depth_features, face_normals[:, :, -1])
return depth_image[..., 0], face_idx >= 0
def render_exact_depth_image(mesh, elev, azim, renderer, size=(1024, 1024)):
depth, mask_bool = render_depth(mesh, elev, azim, renderer)
mask_bool = mask_bool[0]
depth = depth[0]
depth_norm = torch.zeros_like(depth)
fg = depth[mask_bool]
depth_norm[mask_bool] = (fg - fg.min()) / (fg.max() - fg.min() + 1e-8)
depth_rgb = depth_norm.unsqueeze(0).unsqueeze(0).repeat(1, 3, 1, 1)
depth_rgb = torch.nn.functional.interpolate(depth_rgb, size=size, mode="bilinear", align_corners=False)
depth_rgb = depth_rgb[0].clamp(0, 1).permute(1, 2, 0).cpu().numpy()
return Image.fromarray((depth_rgb * 255.0).round().astype(np.uint8))
def _tensor_to_rgb_uint8(image_tensor):
arr = image_tensor[0].detach().cpu().float().permute(1, 2, 0).clamp(0, 1).numpy()
return (arr * 255.0).round().astype(np.uint8)
def save_pixel_overlay(rgb_uint8, pixel, save_path):
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
plt.imshow(rgb_uint8)
plt.scatter([pixel[0]], [pixel[1]], c="red", s=40)
plt.axis("off")
plt.savefig(save_path, bbox_inches="tight", pad_inches=0)
plt.close()
def render_gt_region_with_points(mesh, elev, azim, renderer, render_res, frontview_std,
background, gt_label,
reference_pixel, bsb_vertex, save_path):
bg = torch.tensor(background).to(device)
# Base grey/lit render, plus vertex projections for the point markers.
grey_rgb = torch.tensor(GREY, device=device).repeat(mesh.vertices.shape[0], 1)
setcolor_mesh(mesh, grey_rgb)
base_images, _, _, _, verts_img, _ = renderer.render_views(
mesh, num_views=1, random_views=False,
center_elev=torch.tensor([elev]), center_azim=torch.tensor([azim]),
std=frontview_std, return_views=True, return_coordinates=True,
lighting=True, background=bg)
base_rgb = _tensor_to_rgb_uint8(base_images)
region_prob = torch.zeros(mesh.vertices.shape[0], dtype=torch.float32, device=device)
region_prob[torch.from_numpy(gt_label.astype(bool)).to(device)] = 1.0
prob_tensor = torch.stack([1.0 - region_prob, region_prob], dim=-1).unsqueeze(0)
setcolor_mesh_batched(mesh, prob_tensor)
(mask_images,) = renderer.render_views(
mesh, num_views=1, random_views=False,
center_elev=torch.tensor([elev]), center_azim=torch.tensor([azim]),
std=frontview_std, lighting=False, background=None)
region_mask_2d = mask_images[0, 1].detach().cpu().numpy()
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
plt.figure(figsize=(5, 5))
ax = plt.gca()
ax.imshow(base_rgb)
show_mask(region_mask_2d, ax)
ax.scatter([reference_pixel[0]], [reference_pixel[1]], c="lime", marker="o",
s=100, edgecolors="black", linewidths=1, label="reference vertex")
if bsb_vertex is not None:
verts_pix = to_pixel_coordinates(verts_img, render_res - 1, render_res - 1).round().long()
bsb_x = int(verts_pix[0][bsb_vertex][0].item())
bsb_y_raw = int(verts_pix[0][bsb_vertex][1].item())
bsb_pixel = (bsb_x, render_res - 1 - bsb_y_raw)
ax.scatter([bsb_pixel[0]], [bsb_pixel[1]], c="orange", marker="o",
s=100, edgecolors="black", linewidths=1, label="BSB vertex")
ax.legend(loc="upper right", fontsize=8)
ax.axis("off")
plt.savefig(save_path, bbox_inches="tight", pad_inches=0)
plt.close()
def generate_natural_image(mesh, vertex, generator, prompt, args):
renderer = Renderer(dim=(args.render_res, args.render_res), radius=args.radius)
view = find_visible_view(mesh, vertex, renderer, args.render_res,
args.frontview_std, args.background)
if view is None:
return None
elev, azim, x, y = view
print("[stage1] visible view found: elev=%.4f azim=%.4f pixel=(%d, %d)" % (elev, azim, x, y))
clean = render_clean_view(mesh, elev, azim, renderer, args.frontview_std, args.background)
Image.fromarray(_tensor_to_rgb_uint8(clean)).save(os.path.join(args.out_dir, "render_grey.png"))
depth_estimation = render_exact_depth_image(mesh, elev, azim, renderer)
depth_estimation.save(os.path.join(args.out_dir, "depth_estimation.png"))
# FLUX depth-ControlNet natural image.
image = generator.generate(prompt, depth_estimation,
num_inference_steps=args.num_inference_steps,
guidance_scale=args.guidance_scale)
gen_rgb = _tensor_to_rgb_uint8(image)
gen_rgb = cv2.resize(gen_rgb, (args.render_res, args.render_res), interpolation=cv2.INTER_AREA)
pixel = (int(x), int(y))
gen_path = os.path.join(args.out_dir, "generated.png")
Image.fromarray(gen_rgb).convert("RGB").save(gen_path)
save_pixel_overlay(gen_rgb, pixel, os.path.join(args.out_dir, "generated_with_pixel.png"))
print("[stage1] generated image -> %s ; query pixel=(%d, %d)" % (gen_path, pixel[0], pixel[1]))
return gen_path, pixel, elev, azim
def run_bsb(generated_image_path, pixel, args):
bsb_save_dir = os.path.join(args.out_dir, "bsb")
os.makedirs(bsb_save_dir, exist_ok=True)
# BSB config knobs not exposed as CLI args here (visualization details /
# rarely-changed defaults) -- matches image_shape_correspondence.py's defaults.
bsb_args = argparse.Namespace(
img_path=generated_image_path, object_pixel=list(pixel), test_pixel=list(pixel),
encoder_dino_f_path=args.encoder_dino_f_path, save_dir=bsb_save_dir,
object_mask_idx=args.object_mask_idx, test_mask_idx=args.test_mask_idx,
num_cand_verts=args.num_cand_verts,
batch_size_sim=10, batch_size_iou=100, compute_seg_for_non_bsb=0,
load_object_seg_mask=0, load_pix_dino_features=0, add_test_box=0,
test_box=[100, 150, 480, 650],
show_object_click=1, show_test_click=1, show_test_box=1, show_nn_pix=1,
mask_color=[30, 144, 255, 0.6], mask_color_nn_pix=[205, 255, 5, 0.6],
)
sam = load_img_seg_model(args.sam_dir, args.sam_fname, args.sam_model_type)
image = cv2.cvtColor(cv2.imread(generated_image_path), cv2.COLOR_BGR2RGB)
bsb = BestSegmentationBuddies(bsb_args, sam)
bsb_vert_idx, compute_seg = bsb.find_bsb(image)
bsb_vert_idx = None if bsb_vert_idx is None else int(bsb_vert_idx)
result = {"bsb_vert_idx": bsb_vert_idx, "compute_seg": bool(compute_seg)}
with open(os.path.join(args.out_dir, "bsb_result.json"), 'w') as f:
json.dump(result, f)
print("[stage2] BSB result: %s" % result)
return bsb_vert_idx
def decide_success(bsb_vert_idx, gt_label):
gt_region = gt_label.astype(bool)
if bsb_vert_idx is None:
return False, None
in_region = bool(gt_region[bsb_vert_idx])
return in_region, in_region
def main():
args = parse_args()
os.makedirs(args.out_dir, exist_ok=True)
name = infer_shape_name(args.obj_path)
# Reproducibility for the view search.
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# Ground-truth region on the shape.
gt_label = np.load(args.gt_label_path)
# Stage 1: obtain a natural image + the target vertex's pixel.
mesh = elev = azim = None
if args.generated_image is not None:
if args.object_pixel is None:
raise ValueError("--object_pixel (x y) is required when --generated_image is given.")
vertex = args.vertex
generated_image_path = args.generated_image
pixel = (int(args.object_pixel[0]), int(args.object_pixel[1]))
print("[stage1] skipped; using %s with pixel=(%d, %d)" % (generated_image_path, pixel[0], pixel[1]))
if os.path.isfile(args.obj_path):
mesh = loadmesh(dir=args.obj_path, name=name, load_rings=True)
mesh.face_attributes = mesh.face_attributes.float()
renderer = Renderer(dim=(args.render_res, args.render_res), radius=args.radius)
view = find_visible_view(mesh, vertex, renderer, args.render_res,
args.frontview_std, args.background)
if view is not None:
elev, azim, _, _ = view
else:
print("[viz] no viewpoint sees vertex %d; skipping GT-region visualization." % vertex)
mesh = None
else:
print("[viz] --obj_path %r not found; skipping GT-region visualization." % args.obj_path)
else:
mesh = loadmesh(dir=args.obj_path, name=name, load_rings=True)
mesh.face_attributes = mesh.face_attributes.float()
vertex = args.vertex
prompt = args.prompt if args.prompt is not None else \
"A picture of a %s in a realistic environment" % name
print("[stage1] shape=%s vertex=%d prompt=%r" % (name, vertex, prompt))
generator = DepthControlNetGenerator(
flux_model_path=args.flux_model_path,
device_str=str(device),
seed=args.seed,
deterministic=args.deterministic,
)
result = generate_natural_image(mesh, vertex, generator, prompt, args)
if result is None:
print("[stage1] FAILED: no viewpoint sees vertex %d -> not a success." % vertex)
_write_report(args, name, vertex, None, None, False)
return
generated_image_path, pixel, elev, azim = result
del generator
torch.cuda.empty_cache()
bsb_vert_idx = run_bsb(generated_image_path, pixel, args)
success, in_region = decide_success(bsb_vert_idx, gt_label)
annotated_path = None
if mesh is not None:
annotated_path = os.path.join(args.out_dir, "render_gt_region_with_points.png")
renderer = Renderer(dim=(args.render_res, args.render_res), radius=args.radius)
render_gt_region_with_points(
mesh, elev, azim, renderer, args.render_res, args.frontview_std, args.background,
gt_label, pixel, bsb_vert_idx, annotated_path)
print("[viz] GT region + points -> %s" % annotated_path)
_write_report(args, name, vertex, pixel, bsb_vert_idx, success, annotated_path)
def _write_report(args, name, vertex, pixel, bsb_vert_idx, success, annotated_path=None):
report = {
"shape": name,
"obj_path": args.obj_path,
"gt_label_path": args.gt_label_path,
"target_vertex": vertex,
"query_pixel": None if pixel is None else [int(pixel[0]), int(pixel[1])],
"bsb_vert_idx": None if bsb_vert_idx is None else int(bsb_vert_idx),
"success": bool(success),
"annotated_render_path": annotated_path,
}
report_path = os.path.join(args.out_dir, "success_report.json")
with open(report_path, "w") as f:
json.dump(report, f, indent=2)
print("\n" + "=" * 60)
print(json.dumps(report, indent=2))
print("SUCCESS" if success else "FAILURE")
print("report -> %s" % report_path)
print("=" * 60)
def parse_args():
p = argparse.ArgumentParser(description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
# --- shape / region / vertex ---
p.add_argument('--obj_path', type=str,
default='/net/projects/ranalab/itailang/mesh_feature_fields/partnet_data_v0/bag/12008/12008.obj')
p.add_argument('--gt_label_path', type=str,
default='/net/projects/ranalab/itailang/mesh_feature_fields/partnet_iseg_experiments/bag/12008/decoder/region_1_gt_label.npy',
help='Per-vertex ground-truth labels (.npy). The GT region is where the label is truthy (nonzero).')
p.add_argument('--vertex', type=int, required=True, help='Target vertex index.')
# --- stage 1: generation (FLUX depth-ControlNet) ---
p.add_argument('--flux_model_path', type=str, default='/net/scratch/dwlyu/flux_depth')
p.add_argument('--prompt', type=str, default=None)
p.add_argument('--num_inference_steps', type=int, default=30)
p.add_argument('--guidance_scale', type=float, default=10.0)
p.add_argument('--seed', type=int, default=12)
p.add_argument('--deterministic', action=argparse.BooleanOptionalAction, default=True,
help='If set (default), the FLUX generation step uses --seed, so the same '
'view/prompt/depth map reproduces the same generated image. Pass '
'--no-deterministic to draw a fresh random seed each run instead (the '
'seed actually used is printed either way).')
p.add_argument('--render_res', type=int, default=224)
p.add_argument('--radius', type=float, default=2.0)
p.add_argument('--frontview_std', type=float, default=4.0)
p.add_argument('--background', nargs=3, type=float, default=[1.0, 1.0, 1.0])
# --- stage 1 bypass: reuse a precomputed generated image ---
p.add_argument('--generated_image', type=str, default=None,
help='Skip generation and use this image. Requires --object_pixel.')
p.add_argument('--object_pixel', nargs=2, type=int, default=None,
help='(x y) pixel of the vertex in --generated_image.')
# --- stage 2: BSB ---
p.add_argument('--encoder_dino_f_path', type=str,
default='/net/projects/ranalab/dwlyu/multi_iSeg/evaluation_full_new/bag_12008_12008/dino/pred_f.pth',
help='Per-vertex distilled DINOv2 features (pred_f.pth) for this shape.')
p.add_argument('--object_mask_idx', type=int, default=2)
p.add_argument('--test_mask_idx', type=int, default=0)
p.add_argument('--num_cand_verts', type=int, default=100)
# --- image segmentation (SAM) model ---
p.add_argument('--sam_dir', type=str,
default='/net/projects/ranalab/fei/mesh_feature_fields/SAM/model_checkpoints/')
p.add_argument('--sam_fname', type=str, default='sam_vit_h_4b8939.pth')
p.add_argument('--sam_model_type', type=str, default='vit_h')
# --- output ---
p.add_argument('--out_dir', type=str, default=os.path.join('data', 'eval'))
return p.parse_args()
if __name__ == '__main__':
main()