from torch.utils.data import Dataset import numpy as np import os import random import torchvision.transforms as transforms from PIL import Image, ImageOps import cv2 import torch from PIL.ImageFilter import GaussianBlur import trimesh import logging log = logging.getLogger('trimesh') log.setLevel(40) def load_trimesh(root_dir): folders = os.listdir(root_dir) meshs = {} for i, f in enumerate(folders): sub_name = f meshs[sub_name] = trimesh.load(os.path.join(root_dir, f, '%s_100k.obj' % sub_name)) return meshs def save_samples_truncted_prob(fname, points, prob): ''' Save the visualization of sampling to a ply file. Red points represent positive predictions. Green points represent negative predictions. :param fname: File name to save :param points: [N, 3] array of points :param prob: [N, 1] array of predictions in the range [0~1] :return: ''' r = (prob > 0.5).reshape([-1, 1]) * 255 g = (prob < 0.5).reshape([-1, 1]) * 255 b = np.zeros(r.shape) to_save = np.concatenate([points, r, g, b], axis=-1) return np.savetxt(fname, to_save, fmt='%.6f %.6f %.6f %d %d %d', comments='', header=( 'ply\nformat ascii 1.0\nelement vertex {:d}\nproperty float x\nproperty float y\nproperty float z\nproperty uchar red\nproperty uchar green\nproperty uchar blue\nend_header').format( points.shape[0]) ) class TrainDataset(Dataset): @staticmethod def modify_commandline_options(parser, is_train): return parser def __init__(self, opt, phase='train'): self.opt = opt self.projection_mode = 'orthogonal' # Path setup self.root = self.opt.dataroot self.RENDER = os.path.join(self.root, 'RENDER') self.MASK = os.path.join(self.root, 'MASK') self.PARAM = os.path.join(self.root, 'PARAM') self.UV_MASK = os.path.join(self.root, 'UV_MASK') self.UV_NORMAL = os.path.join(self.root, 'UV_NORMAL') self.UV_RENDER = os.path.join(self.root, 'UV_RENDER') self.UV_POS = os.path.join(self.root, 'UV_POS') self.OBJ = os.path.join(self.root, 'GEO', 'OBJ') self.B_MIN = np.array([-128, -28, -128]) self.B_MAX = np.array([128, 228, 128]) self.is_train = (phase == 'train') self.load_size = self.opt.loadSize self.num_views = self.opt.num_views self.num_sample_inout = self.opt.num_sample_inout self.num_sample_color = self.opt.num_sample_color self.yaw_list = list(range(0,360,1)) self.pitch_list = [0] self.subjects = self.get_subjects() # PIL to tensor self.to_tensor = transforms.Compose([ transforms.Resize(self.load_size), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) ]) # augmentation self.aug_trans = transforms.Compose([ transforms.ColorJitter(brightness=opt.aug_bri, contrast=opt.aug_con, saturation=opt.aug_sat, hue=opt.aug_hue) ]) self.mesh_dic = load_trimesh(self.OBJ) def get_subjects(self): all_subjects = os.listdir(self.RENDER) var_subjects = np.loadtxt(os.path.join(self.root, 'val.txt'), dtype=str) if len(var_subjects) == 0: return all_subjects if self.is_train: return sorted(list(set(all_subjects) - set(var_subjects))) else: return sorted(list(var_subjects)) def __len__(self): return len(self.subjects) * len(self.yaw_list) * len(self.pitch_list) def get_render(self, subject, num_views, yid=0, pid=0, random_sample=False): ''' Return the render data :param subject: subject name :param num_views: how many views to return :param view_id: the first view_id. If None, select a random one. :return: 'img': [num_views, C, W, H] images 'calib': [num_views, 4, 4] calibration matrix 'extrinsic': [num_views, 4, 4] extrinsic matrix 'mask': [num_views, 1, W, H] masks ''' pitch = self.pitch_list[pid] # The ids are an even distribution of num_views around view_id view_ids = [self.yaw_list[(yid + len(self.yaw_list) // num_views * offset) % len(self.yaw_list)] for offset in range(num_views)] if random_sample: view_ids = np.random.choice(self.yaw_list, num_views, replace=False) calib_list = [] render_list = [] mask_list = [] extrinsic_list = [] for vid in view_ids: param_path = os.path.join(self.PARAM, subject, '%d_%d_%02d.npy' % (vid, pitch, 0)) render_path = os.path.join(self.RENDER, subject, '%d_%d_%02d.jpg' % (vid, pitch, 0)) mask_path = os.path.join(self.MASK, subject, '%d_%d_%02d.png' % (vid, pitch, 0)) # loading calibration data param = np.load(param_path, allow_pickle=True) # pixel unit / world unit ortho_ratio = param.item().get('ortho_ratio') # world unit / model unit scale = param.item().get('scale') # camera center world coordinate center = param.item().get('center') # model rotation R = param.item().get('R') translate = -np.matmul(R, center).reshape(3, 1) extrinsic = np.concatenate([R, translate], axis=1) extrinsic = np.concatenate([extrinsic, np.array([0, 0, 0, 1]).reshape(1, 4)], 0) # Match camera space to image pixel space scale_intrinsic = np.identity(4) scale_intrinsic[0, 0] = scale / ortho_ratio scale_intrinsic[1, 1] = -scale / ortho_ratio scale_intrinsic[2, 2] = scale / ortho_ratio # Match image pixel space to image uv space uv_intrinsic = np.identity(4) uv_intrinsic[0, 0] = 1.0 / float(self.opt.loadSize // 2) uv_intrinsic[1, 1] = 1.0 / float(self.opt.loadSize // 2) uv_intrinsic[2, 2] = 1.0 / float(self.opt.loadSize // 2) # Transform under image pixel space trans_intrinsic = np.identity(4) mask = Image.open(mask_path).convert('L') render = Image.open(render_path).convert('RGB') if self.is_train: # Pad images pad_size = int(0.1 * self.load_size) render = ImageOps.expand(render, pad_size, fill=0) mask = ImageOps.expand(mask, pad_size, fill=0) w, h = render.size th, tw = self.load_size, self.load_size # random flip if self.opt.random_flip and np.random.rand() > 0.5: scale_intrinsic[0, 0] *= -1 render = transforms.RandomHorizontalFlip(p=1.0)(render) mask = transforms.RandomHorizontalFlip(p=1.0)(mask) # random scale if self.opt.random_scale: rand_scale = random.uniform(0.9, 1.1) w = int(rand_scale * w) h = int(rand_scale * h) render = render.resize((w, h), Image.BILINEAR) mask = mask.resize((w, h), Image.NEAREST) scale_intrinsic *= rand_scale scale_intrinsic[3, 3] = 1 # random translate in the pixel space if self.opt.random_trans: dx = random.randint(-int(round((w - tw) / 10.)), int(round((w - tw) / 10.))) dy = random.randint(-int(round((h - th) / 10.)), int(round((h - th) / 10.))) else: dx = 0 dy = 0 trans_intrinsic[0, 3] = -dx / float(self.opt.loadSize // 2) trans_intrinsic[1, 3] = -dy / float(self.opt.loadSize // 2) x1 = int(round((w - tw) / 2.)) + dx y1 = int(round((h - th) / 2.)) + dy render = render.crop((x1, y1, x1 + tw, y1 + th)) mask = mask.crop((x1, y1, x1 + tw, y1 + th)) render = self.aug_trans(render) # random blur if self.opt.aug_blur > 0.00001: blur = GaussianBlur(np.random.uniform(0, self.opt.aug_blur)) render = render.filter(blur) intrinsic = np.matmul(trans_intrinsic, np.matmul(uv_intrinsic, scale_intrinsic)) calib = torch.Tensor(np.matmul(intrinsic, extrinsic)).float() extrinsic = torch.Tensor(extrinsic).float() mask = transforms.Resize(self.load_size)(mask) mask = transforms.ToTensor()(mask).float() mask_list.append(mask) render = self.to_tensor(render) render = mask.expand_as(render) * render render_list.append(render) calib_list.append(calib) extrinsic_list.append(extrinsic) return { 'img': torch.stack(render_list, dim=0), 'calib': torch.stack(calib_list, dim=0), 'extrinsic': torch.stack(extrinsic_list, dim=0), 'mask': torch.stack(mask_list, dim=0) } def select_sampling_method(self, subject): if not self.is_train: random.seed(1991) np.random.seed(1991) torch.manual_seed(1991) mesh = self.mesh_dic[subject] surface_points, _ = trimesh.sample.sample_surface(mesh, 4 * self.num_sample_inout) sample_points = surface_points + np.random.normal(scale=self.opt.sigma, size=surface_points.shape) # add random points within image space length = self.B_MAX - self.B_MIN random_points = np.random.rand(self.num_sample_inout // 4, 3) * length + self.B_MIN sample_points = np.concatenate([sample_points, random_points], 0) np.random.shuffle(sample_points) inside = mesh.contains(sample_points) inside_points = sample_points[inside] outside_points = sample_points[np.logical_not(inside)] nin = inside_points.shape[0] inside_points = inside_points[ :self.num_sample_inout // 2] if nin > self.num_sample_inout // 2 else inside_points outside_points = outside_points[ :self.num_sample_inout // 2] if nin > self.num_sample_inout // 2 else outside_points[ :(self.num_sample_inout - nin)] samples = np.concatenate([inside_points, outside_points], 0).T labels = np.concatenate([np.ones((1, inside_points.shape[0])), np.zeros((1, outside_points.shape[0]))], 1) # save_samples_truncted_prob('out.ply', samples.T, labels.T) # exit() samples = torch.Tensor(samples).float() labels = torch.Tensor(labels).float() del mesh return { 'samples': samples, 'labels': labels } def get_color_sampling(self, subject, yid, pid=0): yaw = self.yaw_list[yid] pitch = self.pitch_list[pid] uv_render_path = os.path.join(self.UV_RENDER, subject, '%d_%d_%02d.jpg' % (yaw, pitch, 0)) uv_mask_path = os.path.join(self.UV_MASK, subject, '%02d.png' % (0)) uv_pos_path = os.path.join(self.UV_POS, subject, '%02d.exr' % (0)) uv_normal_path = os.path.join(self.UV_NORMAL, subject, '%02d.png' % (0)) # Segmentation mask for the uv render. # [H, W] bool uv_mask = cv2.imread(uv_mask_path) uv_mask = uv_mask[:, :, 0] != 0 # UV render. each pixel is the color of the point. # [H, W, 3] 0 ~ 1 float uv_render = cv2.imread(uv_render_path) uv_render = cv2.cvtColor(uv_render, cv2.COLOR_BGR2RGB) / 255.0 # Normal render. each pixel is the surface normal of the point. # [H, W, 3] -1 ~ 1 float uv_normal = cv2.imread(uv_normal_path) uv_normal = cv2.cvtColor(uv_normal, cv2.COLOR_BGR2RGB) / 255.0 uv_normal = 2.0 * uv_normal - 1.0 # Position render. each pixel is the xyz coordinates of the point uv_pos = cv2.imread(uv_pos_path, 2 | 4)[:, :, ::-1] ### In these few lines we flattern the masks, positions, and normals uv_mask = uv_mask.reshape((-1)) uv_pos = uv_pos.reshape((-1, 3)) uv_render = uv_render.reshape((-1, 3)) uv_normal = uv_normal.reshape((-1, 3)) surface_points = uv_pos[uv_mask] surface_colors = uv_render[uv_mask] surface_normal = uv_normal[uv_mask] if self.num_sample_color: sample_list = random.sample(range(0, surface_points.shape[0] - 1), self.num_sample_color) surface_points = surface_points[sample_list].T surface_colors = surface_colors[sample_list].T surface_normal = surface_normal[sample_list].T # Samples are around the true surface with an offset normal = torch.Tensor(surface_normal).float() samples = torch.Tensor(surface_points).float() \ + torch.normal(mean=torch.zeros((1, normal.size(1))), std=self.opt.sigma).expand_as(normal) * normal # Normalized to [-1, 1] rgbs_color = 2.0 * torch.Tensor(surface_colors).float() - 1.0 return { 'color_samples': samples, 'rgbs': rgbs_color } def get_item(self, index): # In case of a missing file or IO error, switch to a random sample instead # try: sid = index % len(self.subjects) tmp = index // len(self.subjects) yid = tmp % len(self.yaw_list) pid = tmp // len(self.yaw_list) # name of the subject 'rp_xxxx_xxx' subject = self.subjects[sid] res = { 'name': subject, 'mesh_path': os.path.join(self.OBJ, subject + '.obj'), 'sid': sid, 'yid': yid, 'pid': pid, 'b_min': self.B_MIN, 'b_max': self.B_MAX, } render_data = self.get_render(subject, num_views=self.num_views, yid=yid, pid=pid, random_sample=self.opt.random_multiview) res.update(render_data) if self.opt.num_sample_inout: sample_data = self.select_sampling_method(subject) res.update(sample_data) # img = np.uint8((np.transpose(render_data['img'][0].numpy(), (1, 2, 0)) * 0.5 + 0.5)[:, :, ::-1] * 255.0) # rot = render_data['calib'][0,:3, :3] # trans = render_data['calib'][0,:3, 3:4] # pts = torch.addmm(trans, rot, sample_data['samples'][:, sample_data['labels'][0] > 0.5]) # [3, N] # pts = 0.5 * (pts.numpy().T + 1.0) * render_data['img'].size(2) # for p in pts: # img = cv2.circle(img, (p[0], p[1]), 2, (0,255,0), -1) # cv2.imshow('test', img) # cv2.waitKey(1) if self.num_sample_color: color_data = self.get_color_sampling(subject, yid=yid, pid=pid) res.update(color_data) return res # except Exception as e: # print(e) # return self.get_item(index=random.randint(0, self.__len__() - 1)) def __getitem__(self, index): return self.get_item(index)