Mentions légales du service

Skip to content
Snippets Groups Projects
Commit 79f874a3 authored by KOPANAS Georgios's avatar KOPANAS Georgios
Browse files

simplify cmd line arguments

parent 1e721344
Branches
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
pbnr_pytorch/scene_loaders/ibr_scene.py
+ 58
− 68
View file @ 79f874a3
......@@ -196,7 +196,7 @@ class Camera(nn.Module):
return points_wc, point_features
def render(self, viewpoint_camera, gamma, ewa=False):
def render(self, viewpoint_camera, gamma):
#torch.cuda.synchronize()
#time_begin = time.time()
......@@ -248,77 +248,67 @@ class Camera(nn.Module):
#torch.cuda.synchronize()
#time_begin = time.time()
if ewa:
normal_map = self.normal_map.to("cuda")
normal_map = normal_map.view(3, normal_map.shape[2] * normal_map.shape[3]).permute(1, 0)
normal_map = normal_map / normal_map.norm(2, dim=1, keepdim=True)
filter_pos_x = torch.logical_and(viewspace_points[:, 0] < 1.0, viewspace_points[:, 0] > -1.0)
filter_pos_y = torch.logical_and(viewspace_points[:, 1] < 1.0, viewspace_points[:, 1] > -1.0)
filter_pos = torch.logical_and(filter_pos_x, filter_pos_y)
view_in_ray = (point_cloud - self.camera_center)/(
point_cloud - self.camera_center).norm(dim=1, keepdim=True)
cos_in_view = torch.abs(torch.bmm(normal_map.view(-1, 1, 3), view_in_ray.view(-1, 3, 1)))
view_out_ray = (point_cloud - viewpoint_camera.camera_center) / (
point_cloud - viewpoint_camera.camera_center).norm(dim=1, keepdim=True)
cos_out_view = torch.abs(torch.bmm(normal_map.view(-1, 1, 3), view_out_ray.view(-1, 3, 1)))
filter_slanted = torch.logical_and(cos_in_view > 0.1, cos_out_view > 0.1).squeeze()
filter = torch.logical_and(filter_pos, filter_slanted)
filtered_pc = point_cloud[filter]
filtered_normals = normal_map[filter]
filtered_uncertainty = uncertainty_map_scale_matrix[filter]
if filtered_pc.shape[0] == 0:
return (torch.zeros(1, features.shape[1], viewpoint_camera.image_height, viewpoint_camera.image_width, device="cuda"),
torch.cat([10000.0 * torch.ones((1, viewpoint_camera.image_height, viewpoint_camera.image_width, 102, 1), device="cuda"),
torch.ones((1, viewpoint_camera.image_height, viewpoint_camera.image_width, 102, 1), device="cuda")], dim=-1),
torch.ones((1, viewpoint_camera.image_height, viewpoint_camera.image_width, 1)).cuda(),
torch.zeros(1, 1, viewpoint_camera.image_height, viewpoint_camera.image_width, device="cuda"))
covariance = self.computeCovariance(filtered_pc, filtered_normals, viewpoint_camera)
scaled_covariance = torch.bmm(covariance, filtered_uncertainty)
with torch.no_grad():
eigenval = np.linalg.eigvals(scaled_covariance.cpu())
distortion = eigenval.min(axis=1)/eigenval.max(axis=1)
dist_torch = torch.tensor(distortion).cuda()
ninetinth = np.percentile(eigenval.flatten(), 95)
pixel_sigma = int(np.ceil(3*np.sqrt(ninetinth)))
inv_cov = scaled_covariance.inverse()
#torch.cuda.synchronize()
#time_end = time.time()
#print("Cov: {} | ".format(time_end - time_begin), end="")
#torch.cuda.synchronize()
#time_begin = time.time()
col_image, depth_gmms, num_gmms, mask = rasterizer(viewspace_points[filter],
features[filter],
dist_torch,
inv_cov,
pixel_sigma)
#torch.cuda.synchronize()
#time_end = time.time()
#print("Rast: {} | ".format(time_end - time_begin), end="")
else:
inv_cov = uncertainty_map_scale_matrix*(1.0/(0.66*0.66))
col_image, depth_gmms, num_gmms, mask = rasterizer(viewspace_points,
features,
inv_cov,
self.max_radius)
#torch.cuda.synchronize()
#time_end = time.time()
#print("Rast: {} | ".format(time_end - time_begin), end="")
normal_map = self.normal_map.to("cuda")
normal_map = normal_map.view(3, normal_map.shape[2] * normal_map.shape[3]).permute(1, 0)
normal_map = normal_map / normal_map.norm(2, dim=1, keepdim=True)
filter_pos_x = torch.logical_and(viewspace_points[:, 0] < 1.0, viewspace_points[:, 0] > -1.0)
filter_pos_y = torch.logical_and(viewspace_points[:, 1] < 1.0, viewspace_points[:, 1] > -1.0)
filter_pos = torch.logical_and(filter_pos_x, filter_pos_y)
view_in_ray = (point_cloud - self.camera_center)/(
point_cloud - self.camera_center).norm(dim=1, keepdim=True)
cos_in_view = torch.abs(torch.bmm(normal_map.view(-1, 1, 3), view_in_ray.view(-1, 3, 1)))
view_out_ray = (point_cloud - viewpoint_camera.camera_center) / (
point_cloud - viewpoint_camera.camera_center).norm(dim=1, keepdim=True)
cos_out_view = torch.abs(torch.bmm(normal_map.view(-1, 1, 3), view_out_ray.view(-1, 3, 1)))
filter_slanted = torch.logical_and(cos_in_view > 0.1, cos_out_view > 0.1).squeeze()
filter = torch.logical_and(filter_pos, filter_slanted)
filtered_pc = point_cloud[filter]
filtered_normals = normal_map[filter]
filtered_uncertainty = uncertainty_map_scale_matrix[filter]
if filtered_pc.shape[0] == 0:
return (torch.zeros(1, features.shape[1], viewpoint_camera.image_height, viewpoint_camera.image_width, device="cuda"),
torch.cat([10000.0 * torch.ones((1, viewpoint_camera.image_height, viewpoint_camera.image_width, 102, 1), device="cuda"),
torch.ones((1, viewpoint_camera.image_height, viewpoint_camera.image_width, 102, 1), device="cuda")], dim=-1),
torch.ones((1, viewpoint_camera.image_height, viewpoint_camera.image_width, 1)).cuda(),
torch.zeros(1, 1, viewpoint_camera.image_height, viewpoint_camera.image_width, device="cuda"))
covariance = self.computeCovariance(filtered_pc, filtered_normals, viewpoint_camera)
scaled_covariance = torch.bmm(covariance, filtered_uncertainty)
with torch.no_grad():
eigenval = np.linalg.eigvals(scaled_covariance.cpu())
distortion = eigenval.min(axis=1)/eigenval.max(axis=1)
dist_torch = torch.tensor(distortion).cuda()
ninetinth = np.percentile(eigenval.flatten(), 95)
pixel_sigma = int(np.ceil(3*np.sqrt(ninetinth)))
inv_cov = scaled_covariance.inverse()
#torch.cuda.synchronize()
#time_end = time.time()
#print("Cov: {} | ".format(time_end - time_begin), end="")
#torch.cuda.synchronize()
#time_begin = time.time()
col_image, depth_gmms, num_gmms, mask = rasterizer(viewspace_points[filter],
features[filter],
dist_torch,
inv_cov,
pixel_sigma)
#torch.cuda.synchronize()
#time_end = time.time()
#print("Rast: {} | ".format(time_end - time_begin), end="")
return col_image, depth_gmms, num_gmms, mask
def render_patch(self, viewpoint_camera, patch_size_y, patch_size_x, patch_origin_x, patch_origin_y,
gamma, ewa=False):
gamma):
pixel_size_y = (2 * math.tan(viewpoint_camera.FoVy / 2)) / viewpoint_camera.image_height
pixel_size_x = (2 * math.tan(viewpoint_camera.FoVx / 2)) / viewpoint_camera.image_width
cx = viewpoint_camera.image_width/2
......@@ -334,7 +324,7 @@ class Camera(nn.Module):
zfar=viewpoint_camera.zfar,
image_height=patch_size_y,
image_width=patch_size_x)
return self.render(viewpoint_camera=camera_for_patch, gamma=gamma, ewa=ewa)
return self.render(viewpoint_camera=camera_for_patch, gamma=gamma)
class Scene():
......
pbnr_pytorch/test_path.py
+ 5
− 6
View file @ 79f874a3
......@@ -11,7 +11,7 @@ import json
import numpy as np
import math
def render_viewpoint2(viewpoint_camera, pcloud_cameras, patch=None, ewa=False, gamma=1.0):
def render_viewpoint2(viewpoint_camera, pcloud_cameras, patch=None, gamma=1.0):
if patch==None:
patch = (0, 0, viewpoint_camera.image_width, viewpoint_camera.image_height)
patch_origin_x, patch_origin_y, patch_size_x, patch_size_y = patch
......@@ -25,7 +25,7 @@ def render_viewpoint2(viewpoint_camera, pcloud_cameras, patch=None, ewa=False, g
rendered_point_cloud, depth_gmms, num_gmms, blend_scores = pcloud_cam.render_patch(viewpoint_camera=viewpoint_camera,
patch_origin_x=patch_origin_x, patch_size_x=patch_size_x,
patch_origin_y=patch_origin_y, patch_size_y=patch_size_y,
gamma=gamma, ewa=ewa)
gamma=gamma)
color_stack = torch.cat((color_stack, rendered_point_cloud), dim=0)
blend_scores_stack = torch.cat((blend_scores_stack, blend_scores), dim=0)
features_stack = torch.cat((features_stack, rendered_point_cloud), dim=0)
......@@ -40,7 +40,7 @@ def render_viewpoint2(viewpoint_camera, pcloud_cameras, patch=None, ewa=False, g
image = neural_renderer(features_stack, prob_map * blend_scores_stack)
return image, color_stack, l2_stack.mean()
def render_viewpoint(viewpoint_camera, pcloud_cameras, patch=None, ewa=False, gamma=1.0):
def render_viewpoint(viewpoint_camera, pcloud_cameras, patch=None, gamma=1.0):
print("render_viewpoint")
print(viewpoint_camera.projection_matrix)
features_stack = torch.tensor([]).to(device)
......@@ -51,7 +51,7 @@ def render_viewpoint(viewpoint_camera, pcloud_cameras, patch=None, ewa=False, ga
blend_scores_stack = torch.tensor([]).to(device)
for idx, pcloud_cam in enumerate(pcloud_cameras):
rendered_point_cloud, depth_gmms, num_gmms, blend_scores = pcloud_cam.render(viewpoint_camera=viewpoint_camera,
gamma=gamma, ewa=ewa)
gamma=gamma)
color_stack = torch.cat((color_stack, rendered_point_cloud), dim=0)
blend_scores_stack = torch.cat((blend_scores_stack, blend_scores), dim=0)
features_stack = torch.cat((features_stack, rendered_point_cloud), dim=0)
......@@ -121,7 +121,6 @@ parser.add_argument('--load_iter', required=False, type=int, default=None)
parser.add_argument('--max_radius', required=False, type=int, default=8)
parser.add_argument('--test_cameras', required=False, default=3)
parser.add_argument('--extra_features', type=int, default=6)
parser.add_argument('--sample_views', action='store_true', dest='sample_views')
args = parser.parse_args()
......@@ -150,7 +149,7 @@ with torch.no_grad():
train_viewpoint = scene.getAllTrainCameras()[2]
train_pcloud_cams = scene.getPCloudCamsForScore(train_viewpoint, args.sample_views)
train_pcloud_cams = scene.getPCloudCamsForScore(train_viewpoint, False)
image, image_stack, _ = render_viewpoint(cameras_path[0], train_pcloud_cams, None, True)
print("path")
print(cameras_path[0].world_view_transform)
......
pbnr_pytorch/train_full_pipeline.py
+ 8
− 10
View file @ 79f874a3
......@@ -30,7 +30,7 @@ class F:
sys.stdout = F()
def render_viewpoint(viewpoint_camera, pcloud_cameras, patch=None, ewa=False, gamma=1.0):
def render_viewpoint(viewpoint_camera, pcloud_cameras, patch=None, gamma=1.0):
if patch==None:
patch = (0, 0, viewpoint_camera.image_width, viewpoint_camera.image_height)
patch_origin_x, patch_origin_y, patch_size_x, patch_size_y = patch
......@@ -44,7 +44,7 @@ def render_viewpoint(viewpoint_camera, pcloud_cameras, patch=None, ewa=False, ga
rendered_point_cloud, depth_gmms, num_gmms, blend_scores = pcloud_cam.render_patch(viewpoint_camera=viewpoint_camera,
patch_origin_x=patch_origin_x, patch_size_x=patch_size_x,
patch_origin_y=patch_origin_y, patch_size_y=patch_size_y,
gamma=gamma, ewa=ewa)
gamma=gamma)
color_stack = torch.cat((color_stack, rendered_point_cloud), dim=0)
blend_scores_stack = torch.cat((blend_scores_stack, blend_scores), dim=0)
features_stack = torch.cat((features_stack, rendered_point_cloud), dim=0)
......@@ -86,10 +86,8 @@ parser.add_argument('--skip_validation', action='store_true', dest='skip_validat
parser.add_argument('--extra_features', type=int, default=6)
parser.add_argument('--max_radius', required=False, type=int, default=8)
parser.add_argument('--sample_views', action='store_true', dest='sample_views')
parser.add_argument('--test_cameras', required=False, default=3)
parser.add_argument('--ewa', action='store_true', dest='ewa')
parser.add_argument('--input_views', required=False, default=9)
args = parser.parse_args()
......@@ -161,7 +159,7 @@ while True:
viewpoint_idx = randint(0, len(viewpoint_stack)-1)
viewpoint_cam = viewpoint_stack.pop(viewpoint_idx)
pcloud_cams = scene.getPCloudCamsForScore(viewpoint_cam, args.sample_views)
pcloud_cams = scene.getPCloudCamsForScore(viewpoint_cam, sample=True)
random.shuffle(pcloud_cams)
# Zero out the gradients
......@@ -176,7 +174,7 @@ while True:
patch_size = 150
patch = (randint(0, viewpoint_cam.image_width - patch_size), randint(0, viewpoint_cam.image_height - patch_size), patch_size, patch_size)
image, image_stack, l2 = render_viewpoint(viewpoint_cam, pcloud_cams, patch, args.ewa)
image, image_stack, l2 = render_viewpoint(viewpoint_cam, pcloud_cams, patch)
gt_image = viewpoint_cam.getImageHarmonized().to(device)
gt_image = crop_image(gt_image, patch)
......@@ -218,9 +216,9 @@ while True:
for test_viewpoint in scene.getAllTestCameras():
torch.cuda.empty_cache()
test_pcloud_cams = scene.getPCloudCamsForScore(test_viewpoint, args.sample_views)
test_pcloud_cams = scene.getPCloudCamsForScore(test_viewpoint, sample=False)
random.shuffle(test_pcloud_cams)
image, image_stack, _ = render_viewpoint(test_viewpoint, test_pcloud_cams, None, args.ewa)
image, image_stack, _ = render_viewpoint(test_viewpoint, test_pcloud_cams, None)
gt_image = test_viewpoint.getImageHarmonized().to("cuda")
Ll1 = l1_loss(image, gt_image)
......@@ -253,9 +251,9 @@ while True:
for train_viewpoint in validation_cameras:
torch.cuda.empty_cache()
train_pcloud_cams = scene.getPCloudCamsForScore(train_viewpoint, args.sample_views)
train_pcloud_cams = scene.getPCloudCamsForScore(train_viewpoint, sample=False)
random.shuffle(train_pcloud_cams)
image, image_stack, _ = render_viewpoint(train_viewpoint, train_pcloud_cams, None, args.ewa)
image, image_stack, _ = render_viewpoint(train_viewpoint, train_pcloud_cams, None)
gt_image = train_viewpoint.getImageHarmonized().to("cuda")
Ll1 = l1_loss(image, gt_image)
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment