import argparse import trimesh import numpy as np import os import shutil import glob import sys from multiprocessing import Pool from functools import partial sys.path.append(os.getcwd()) from lib.utils.libmesh import check_mesh_contains from smpl_torch_batch import SMPLModel parser = argparse.ArgumentParser('Sample a watertight mesh.') parser.add_argument('in_folder', type=str, help='Path to input watertight meshes.') parser.add_argument('--out_folder', type=str, help='Path to save the outputs.') parser.add_argument('--ext', type=str, default='obj', help='Extensions for meshes.') parser.add_argument('--n_proc', type=int, default=0, help='Number of processes to use.') parser.add_argument('--resize', action='store_true', help='When active, resizes the mesh to bounding box.') parser.add_argument('--bbox_padding', type=float, default=0., help='Padding for bounding box') parser.add_argument('--pointcloud_folder', type=str, default='pcl_seq', help='Output path for point cloud.') parser.add_argument('--pointcloud_size', type=int, default=100000, help='Size of point cloud.') parser.add_argument('--points_folder', type=str, default='points_seq', help='Output path for points.') parser.add_argument('--points_size', type=int, default=100000, help='Size of points.') parser.add_argument('--points_uniform_ratio', type=float, default=0.5, help='Ratio of points to sample uniformly' 'in bounding box.') parser.add_argument('--points_sigma', type=float, default=0.01, help='Standard deviation of gaussian noise added to points' 'samples on the surfaces.') parser.add_argument('--points_padding', type=float, default=0.1, help='Additional padding applied to the uniformly' 'sampled points on both sides (in total).') parser.add_argument('--overwrite', action='store_true', help='Whether to overwrite output.') parser.add_argument('--float16', action='store_true', help='Whether to use half precision.') parser.add_argument('--packbits', action='store_true', help='Whether to save truth values as bit array.') def main(args): hids = os.listdir(os.path.join(args.in_folder)) seq_folders = [] for hid in hids: seq_folders.extend(glob.glob(os.path.join(args.in_folder, hid, '*'))) seq_folders.sort() print('Total number of sequences: ', len(seq_folders)) if args.n_proc != 0: with Pool(args.n_proc) as p: p.map(partial(process_path, args=args), seq_folders) else: for p in seq_folders: process_path(p, args) def process_path(in_path, args): smpl_model = SMPLModel(model_path='data/human_dataset/smpl_models/model_300_m.pkl') smpl_faces = smpl_model.faces identity, motion = in_path.split('/')[-2:] model_file = os.path.join(in_path, 'smpl_vers.npy') # Export various modalities if args.pointcloud_folder is not None: export_pointcloud(identity, motion, model_file, smpl_faces, args) if args.points_folder is not None: export_points(identity, motion, model_file, smpl_faces, args) print(in_path) def get_loc_scale(mesh, args): # Determine bounding box if not args.resize: # Standard bounding boux loc = np.zeros(3) scale = 1. else: bbox = mesh.bounding_box.bounds # Compute location and scale loc = (bbox[0] + bbox[1]) / 2 scale = (bbox[1] - bbox[0]).max() / (1 - args.bbox_padding) return loc, scale # Export functions def export_pointcloud(identity, motion, model_files, smpl_faces, args): out_folder = os.path.join(args.out_folder, 'D-FAUST', identity, motion, args.pointcloud_folder) if os.path.exists(out_folder): if not args.overwrite: print('Pointcloud already exist: %s' % out_folder) return else: shutil.rmtree(out_folder) # Create out_folder os.makedirs(out_folder) all_vers = np.load(model_files) mesh = trimesh.Trimesh(all_vers[0].squeeze(), smpl_faces.squeeze(), process=False) _, face_idx = mesh.sample(args.pointcloud_size, return_index=True) alpha = np.random.dirichlet((1,)*3, args.pointcloud_size) for it, verts in enumerate(all_vers): out_file = os.path.join(out_folder, '%08d.npz' % it) mesh = trimesh.Trimesh(verts.squeeze(), smpl_faces.squeeze(), process=False) loc = np.zeros(3) scale = np.array([1.]) vertices = mesh.vertices faces = mesh.faces v = vertices[faces[face_idx]] points = (alpha[:, :, None] * v).sum(axis=1) print('Writing pointcloud: %s' % out_file) # Compress if args.float16: dtype = np.float16 else: dtype = np.float32 points = points.astype(dtype) loc = loc.astype(dtype) scale = scale.astype(dtype) np.savez(out_file, points=points, loc=loc, scale=scale) def export_points(identity, motion, model_files, smpl_faces, args): out_folder = os.path.join(args.out_folder, 'D-FAUST', identity, motion, args.points_folder) if os.path.exists(out_folder): if not args.overwrite: print('Points already exist: %s' % out_folder) return else: shutil.rmtree(out_folder) # Create out_folder os.makedirs(out_folder) all_vers = np.load(model_files) n_points_uniform = int(args.points_size * args.points_uniform_ratio) n_points_surface = args.points_size - n_points_uniform for it, verts in enumerate(all_vers): out_file = os.path.join(out_folder, '%08d.npz' % it) mesh = trimesh.Trimesh(verts.squeeze(), smpl_faces.squeeze(), process=False) if not mesh.is_watertight: print('Warning: mesh %s is not watertight!') loc_self, scale_self = get_loc_scale(mesh, args) loc_global = np.array([-0.005493, -0.1888, 0.07587]) scale_global = np.array([2.338]) mesh.apply_translation(-loc_global) mesh.apply_scale(1/scale_global) boxsize = 1 + args.points_padding points_uniform = np.random.rand(n_points_uniform, 3) points_uniform = boxsize * (points_uniform - 0.5) points_uniform = (loc_self + scale_self * points_uniform - loc_global) / scale_global points_surface = mesh.sample(n_points_surface) points_surface += args.points_sigma * \ np.random.randn(n_points_surface, 3) points = np.concatenate([points_uniform, points_surface], axis=0) occupancies = check_mesh_contains(mesh, points) print('Writing points: %s' % out_file) # Compress if args.float16: dtype = np.float16 else: dtype = np.float32 points = points.astype(dtype) loc = loc_global.astype(dtype) scale = scale_global.astype(dtype) if args.packbits: occupancies = np.packbits(occupancies) np.savez(out_file, points=points, occupancies=occupancies, loc=loc, scale=scale) if __name__ == '__main__': args = parser.parse_args() main(args)