model = DeepVoxels(lifting_img_dims=proj_image_dims,
frustrum_img_dims=proj_image_dims,
grid_dims=grid_dims,
use_occlusion_net=not opt.no_occlusion_net,
num_grid_feats=opt.num_grid_feats,
nf0=opt.nf0,
img_sidelength=input_image_dims[0])
model.to(device)
# Projection module
projection = ProjectionHelper(projection_intrinsic=proj_intrinsic,
lifting_intrinsic=lift_intrinsic,
depth_min=depth_min,
depth_max=depth_max,
projection_image_dims=proj_image_dims,
lifting_image_dims=proj_image_dims,
grid_dims=grid_dims,
voxel_size=voxel_size,
device=device,
frustrum_depth=frustrum_depth,
near_plane=near_plane)
# L1 loss
criterionL1 = nn.L1Loss(reduction='mean').to(device)
# GAN loss
criterionGAN = GANLoss().to(device)
discriminator = PatchDiscriminator(input_nc=3).to(device)
# Optimizers
optimizerD = torch.optim.Adam(discriminator.parameters(), lr=opt.lr)
proj_image_dims = [68, 50]
# TODO: load intrinsic from txt
intrinsic = util.make_intrinsic(opt.fx, opt.fy, opt.mx, opt.my)
intrinsic = util.adjust_intrinsic(intrinsic, [opt.intrinsic_image_width, opt.intrinsic_image_height], proj_image_dims)
intrinsic = intrinsic.cuda()
grid_dims = [opt.grid_dimX, opt.grid_dimY, opt.grid_dimZ]
column_height = opt.grid_dimZ
batch_size = opt.batch_size
num_images = opt.num_nearest_images
grid_centerX = opt.grid_dimX // 2
grid_centerY = opt.grid_dimY // 2
color_mean = ENET_TYPES[opt.model2d_type][1]
color_std = ENET_TYPES[opt.model2d_type][2]
num_classes = opt.num_classes
projection = ProjectionHelper(intrinsic, opt.depth_min, opt.depth_max, proj_image_dims, grid_dims, opt.voxel_size)
def load_frames_multi(data_path, frame_indices, depth_images, color_images, poses, color_mean, color_std):
num_images = 1
depth_files = [os.path.join(data_path, 'depth20.png')]
color_files = [os.path.join(data_path, 'color20.jpg')]
pose_files = [os.path.join(data_path, 'pose20.txt')]
# TODO: change batch_size
# batch_size = frame_indices.size(0) * num_images
batch_size = 1
depth_image_dims = [depth_images.shape[2], depth_images.shape[1]]
color_image_dims = [color_images.shape[3], color_images.shape[2]]
normalize = transforms.Normalize(mean=color_mean, std=color_std)
# load data
mesh_vertices = np.delete(mesh_vertices, 3, 1)
pts = np.ones((mesh_vertices.shape[0], 4))
pts[:, 0:3] = mesh_vertices[:, 0:3]
pts = np.dot(pts, axis_align_matrix.transpose()) # Nx4
mesh_vertices[:, 0:3] = pts[:, 0:3]
mesh_vertices = mesh_vertices.astype('float32')
return mesh_vertices
if __name__ == '__main__':
for scene_id in os.listdir(scene_path):
intrinsic_str = read_lines_from_file(rawdata_path + scene_id +
'/intrinsic_depth.txt')
fx = float(intrinsic_str[0].split()[0])
fy = float(intrinsic_str[1].split()[1])
mx = float(intrinsic_str[0].split()[2])
my = float(intrinsic_str[1].split()[2])
intrinsic = util.make_intrinsic(fx, fy, mx, my)
intrinsic = util.adjust_intrinsic(
intrinsic, [opt.intrinsic_image_width, opt.intrinsic_image_height],
proj_image_dims)
intrinsic = intrinsic.cuda()
projection = ProjectionHelper(intrinsic, opt.depth_min, opt.depth_max,
proj_image_dims)
vertices = scannet_projection(intrinsic, projection, scene_id)
save_path = "/home/extra/extra/outputs/"
if not os.path.exists(save_path):
os.makedirs(save_path)
np.save(save_path + scene_id + "_features.npy", vertices)
print("finished")
