import yaml import pickle from subprocess import run import numpy as np import torch import torch.nn.functional as F # import smplx import shutil import os from pathlib import Path import trimesh import pyrender import math import matplotlib.pyplot as plt import cv2 from contextlib import contextmanager from PIL import Image from scipy.spatial.transform import Rotation def project_keypoints_to_2d(keyp_3d, camera_params, img_size=1024): """ Project 3D keypoints to 2D image coordinates using weak perspective camera. Applies the same transformation as the mesh renderer (180 degree rotation around X-axis). Uses the same projection matrix as WeakPerspectiveCamera. Args: keyp_3d: (N, 3) or (1, N, 3) array of 3D keypoints camera_params: [sx, sy, tx, ty] camera parameters img_size: image size (assumes square image) Returns: keyp_2d: (N, 2) array of 2D keypoints in image coordinates """ if keyp_3d.ndim == 3: keyp_3d = keyp_3d[0] # (N, 3) # Apply the same 180-degree rotation around X-axis as in the renderer # This matches the transformation in _create_mesh_from_trimesh: # transform = trimesh.transformations.rotation_matrix(math.radians(180), [1, 0, 0]) rot_matrix = np.array([ [1, 0, 0], [0, -1, 0], # cos(180) = -1, sin(180) = 0 [0, 0, -1] ]) keyp_3d_transformed = keyp_3d @ rot_matrix.T # Apply rotation sx, sy, tx, ty = camera_params # Apply projection matrix as in WeakPerspectiveCamera.get_projection_matrix() # P[0, 0] = sx, P[0, 3] = tx * sx # P[1, 1] = sy, P[1, 3] = -ty * sy <- Note the negative sign! x_proj = sx * keyp_3d_transformed[:, 0] + tx * sx y_proj = sy * keyp_3d_transformed[:, 1] - ty * sy # Negative ty! # Convert from NDC (normalized device coordinates) [-1, 1] to pixel coordinates [0, img_size] x_pixel = (x_proj + 1) * img_size / 2 y_pixel = (1 - y_proj) * img_size / 2 # Flip Y for image coordinates keyp_2d = np.stack([x_pixel, y_pixel], axis=1) return keyp_2d def draw_keypoints_on_image(image, keyp_2d, radius=5, color=(255, 0, 0), thickness=-1): """ Draw keypoints on image. Args: image: (H, W, 3) or (H, W, 4) numpy array keyp_2d: (N, 2) array of 2D keypoint coordinates radius: circle radius color: RGB color tuple thickness: circle thickness (-1 for filled) Returns: image with keypoints drawn """ img = np.array(image).copy() for i, (x, y) in enumerate(keyp_2d): if 0 <= x < img.shape[1] and 0 <= y < img.shape[0]: cv2.circle(img, (int(x), int(y)), radius, color, thickness) # Optional: add keypoint index # cv2.putText(img, str(i), (int(x)+5, int(y)+5), # cv2.FONT_HERSHEY_SIMPLEX, 0.3, color, 1) return img def show_progress_bar(current, total, description="Processing"): if total == 0: return progress = (current + 1) / total bar_length = 50 filled_length = int(bar_length * progress) bar = '#' * filled_length + '-' * (bar_length - filled_length) percentage = progress * 100 print(f"\r{description}: [{bar}] {percentage:.1f}% ({current + 1}/{total})", end="", flush=True) if current + 1 == total: print() def copy_with_progress_bar(source_path, dest_path, description="Copying"): if not os.path.exists(source_path): print(f"Warning: {description} source not found at {source_path}") return False # Remove destination if it exists if os.path.exists(dest_path): shutil.rmtree(dest_path) # Create destination directory os.makedirs(dest_path, exist_ok=True) if os.path.isdir(source_path): # Copy directory contents items = list(Path(source_path).rglob("*")) files = [item for item in items if item.is_file()] print(f"{description} {len(files)} files from {source_path} to {dest_path}") for i, file_path in enumerate(files): rel_path = file_path.relative_to(source_path) dest_file = Path(dest_path) / rel_path dest_file.parent.mkdir(parents=True, exist_ok=True) shutil.copy2(file_path, dest_file) # Update progress bar show_progress_bar(i, len(files), description) return True else: # Copy single file print(f"{description} single file from {source_path} to {dest_path}") shutil.copy2(source_path, dest_path) return True def rsync(src, dst): run(["rsync", "-a", src, dst], check=True) def load_yaml(path): with open(path, 'r') as f: return yaml.safe_load(f) def load_pkl(path): with open(path, 'rb') as file: return pickle.load(file) def pitch(): # Positive: down theta = np.deg2rad(np.random.uniform(-10, 100)) rotation_matrix = np.array([ [1, 0, 0], [0, np.cos(theta), -np.sin(theta)], [0, np.sin(theta), np.cos(theta)] ], dtype=np.float32) return rotation_matrix, theta > np.deg2rad(50) def yaw(): # Positive: clockwise theta = np.deg2rad(np.random.uniform(-180, 180)) rotation_matrix = np.array([ [np.cos(theta), 0, -np.sin(-theta)], [0, 1, 0], [np.sin(-theta), 0, np.cos(theta)] ], dtype=np.float32) return rotation_matrix def roll(theta): # Positive: clockwise return np.array([ [1, 0, 0], [0, np.cos(theta), -np.sin(theta)], [0, np.sin(theta), np.cos(theta)] ], dtype=np.float32) def rot6d_to_rotmat(rot6d): assert rot6d.ndim == 2 rot6d = rot6d.view(-1, 3, 2) a1 = rot6d[:, :, 0] a2 = rot6d[:, :, 1] b1 = F.normalize(a1) b2 = F.normalize(a2 - torch.einsum("bi,bi->b", b1, a2).unsqueeze(-1) * b1) b3 = torch.linalg.cross(b1, b2) rotmat = torch.stack((b1, b2, b3), dim=-1) return rotmat def create_pose_rotmat(pose, orient): orient = rot6d_to_rotmat(torch.tensor(orient).reshape(-1, 6)).reshape(1, 3, 3).numpy() pose = rot6d_to_rotmat(torch.tensor(pose).reshape(-1, 6)).reshape(-1, 3, 3).numpy() return np.concatenate([orient, pose], axis=0) def get_bbox(proj): min_x, min_y = np.min(proj, axis=0) max_x, max_y = np.max(proj, axis=0) return (min_x, min_y, max_x, max_y) # class SMALLayer(smplx.SMPLLayer): # NUM_JOINTS = 34 # NUM_BODY_JOINTS = 34 # SHAPE_SPACE_DIM = 145 # def __init__(self, *args, **kwargs): # super().__init__(*args, **kwargs) # self.vertex_joint_selector.extra_joints_idxs = torch.empty(0, dtype=torch.int32) def weak_perspective_project(verts, scale, tx, ty): proj = verts[:, :2] * scale proj[:, 0] += tx proj[:, 1] += ty return proj def convert_to_pixel_coords(scale, tx, ty, resolution=1024): tx_px = (tx + 0.5) * resolution ty_px = (ty + 0.5) * resolution s_wp = scale * resolution return s_wp, tx_px, ty_px def save_vertices_obj(vertices, faces, save_path): with open(save_path, 'w') as f: for v in vertices: f.write(f'v {v[0]} {v[1]} {v[2]}\n') for face in faces + 1: f.write(f'f {face[0]} {face[1]} {face[2]}\n') def deduce_weak_perspective_params(verts, img_size=(1024, 1024), random=False): """ Compute weak perspective camera parameters to fit mesh in image. Args: verts: (N, 3) vertices img_size: (width, height) image size random: if True, add randomness to scale and translation; if False, fit exactly Returns: scale, tx, ty: camera parameters """ min_x, min_y = verts[:, :2].min(axis=0) max_x, max_y = verts[:, :2].max(axis=0) scale_x = img_size[0] / (max_x - min_x) scale_y = img_size[1] / (max_y - min_y) scale = min(scale_x, scale_y) scale *= 0.9 if random: scale *= np.random.uniform(0.9, 1.0) scaled_width = scale * (max_x - min_x) scaled_height = scale * (max_y - min_y) tx_min = 0 - scale * min_x tx_max = img_size[0] - scale * max_x ty_min = 0 - scale * min_y ty_max = img_size[1] - scale * max_y if random: tx = np.random.uniform(tx_min, tx_max) ty_mid = (ty_min + ty_max) / 2 ty = np.random.uniform(ty_mid, ty_max) else: # Center the mesh perfectly tx = (tx_min + tx_max) / 2 ty = (ty_min + ty_max) / 2 return scale, tx, ty class WeakPerspectiveCamera(pyrender.Camera): def __init__(self, scale, translation, znear=10.0, zfar=1000.0): super().__init__(znear=znear, zfar=zfar) self.scale = np.asarray(scale, dtype=float).ravel()[:2] self.translation = np.asarray(translation, dtype=float).ravel()[:2] def get_projection_matrix(self, width=None, height=None): P = np.eye(4) sx, sy = self.scale tx, ty = self.translation P[0, 0] = sx P[1, 1] = sy P[0, 3] = tx * sx P[1, 3] = -ty * sy P[2, 2] = -0.1 return P class MeshRenderer: def __init__(self, faces=None, resolution=(1024, 1024), randomize_light_orientation=False): self.faces = faces self.resolution = resolution self.randomize_light_orientation = randomize_light_orientation self.renderer = pyrender.OffscreenRenderer(*resolution) self.scene = pyrender.Scene( bg_color=[0, 0, 0, 0], ambient_light=(0.5, 0.5, 0.5) ) self.light_nodes = [] self._setup_lights(randomize_orientation=randomize_light_orientation) def _setup_lights(self, randomize_orientation=False): # Remove existing lights first for node in self.light_nodes: self.scene.remove_node(node) self.light_nodes = [] # Base directions for directional lights (pointing towards origin) # DirectionalLight shines along its local -Z axis, so we create rotations # that point the light from different directions base_directions = np.array([ [0, -1, 1], # from above-front [0, 1, 1], # from above-back [1, 1, 2], # from above-right-back ], dtype=np.float64) # Normalize to get unit direction vectors base_directions = base_directions / np.linalg.norm(base_directions, axis=1, keepdims=True) if randomize_orientation: # Apply random 3D rotation to all light directions random_rotation = Rotation.random() directions = random_rotation.apply(base_directions) else: directions = base_directions light = pyrender.DirectionalLight(color=[1.0, 1.0, 1.0], intensity=0.8) for direction in directions: # Create rotation matrix that aligns -Z axis with the light direction # Light shines along -Z in its local frame, so we want -Z to point along 'direction' # This means Z should point along '-direction' z_axis = -direction # The light's Z axis (opposite of light direction) # Create orthonormal basis # Pick an arbitrary vector not parallel to z_axis for cross product up = np.array([0, 1, 0]) if abs(z_axis[1]) < 0.9 else np.array([1, 0, 0]) x_axis = np.cross(up, z_axis) x_axis = x_axis / np.linalg.norm(x_axis) y_axis = np.cross(z_axis, x_axis) # Build rotation matrix (columns are the local axes in world coordinates) pose = np.eye(4) pose[:3, 0] = x_axis pose[:3, 1] = y_axis pose[:3, 2] = z_axis node = self.scene.add(light, pose=pose) self.light_nodes.append(node) def _calculate_surface_normals(self, vertices, faces): normals = np.cross( vertices[faces[:, 1]] - vertices[faces[:, 0]], vertices[faces[:, 2]] - vertices[faces[:, 1]] ) normals /= np.linalg.norm(normals, axis=1)[:, None] return (normals + 1) / 2 * 255 def _create_mesh_from_vertices(self, vertices, color=None): """Create mesh from vertices and faces (legacy method)""" mesh = trimesh.Trimesh(vertices=vertices, faces=self.faces, process=False) mesh = mesh.subdivide_loop(iterations=2) transform = trimesh.transformations.rotation_matrix(math.radians(180), [1, 0, 0]) mesh.apply_transform(transform) if color is not None: material = pyrender.MetallicRoughnessMaterial( baseColorFactor=[*color, 1.0], metallicFactor=0.1, alphaMode="OPAQUE" ) return pyrender.Mesh.from_trimesh(mesh, material=material, smooth=True) else: normals = self._calculate_surface_normals(mesh.vertices, mesh.faces) mesh.visual.face_colors = normals.astype(np.uint8) return pyrender.Mesh.from_trimesh(mesh, material=None, smooth=False) def _create_mesh_from_trimesh(self, trimesh_obj, color=None): """Create mesh from a trimesh object (supports textures)""" mesh = trimesh_obj.copy() # Check if mesh has texture - if so, don't subdivide as it breaks UV coordinates has_texture = hasattr(mesh.visual, 'uv') and mesh.visual.uv is not None if not has_texture: mesh = mesh.subdivide_loop(iterations=2) transform = trimesh.transformations.rotation_matrix(math.radians(180), [1, 0, 0]) mesh.apply_transform(transform) if color is not None: # Override with solid color material = pyrender.MetallicRoughnessMaterial( baseColorFactor=[*color, 1.0], metallicFactor=0.1, alphaMode="OPAQUE" ) return pyrender.Mesh.from_trimesh(mesh, material=material, smooth=True) else: # Use existing visual (texture or vertex colors) # pyrender automatically handles TextureVisuals from trimesh return pyrender.Mesh.from_trimesh(mesh, smooth=True) def render(self, mesh_or_vertices, camera_params, color=None, depth_only=False): """ Render a mesh with camera parameters. Args: mesh_or_vertices: Either a trimesh.Trimesh object or numpy array of vertices camera_params: Camera parameters [sx, sy, tx, ty] color: Optional color override (R, G, B) in [0, 1] depth_only: If True, only return depth map """ # Randomize light orientation on each render call if enabled if self.randomize_light_orientation: self._setup_lights(randomize_orientation=True) # Determine if input is a trimesh object or vertices array if isinstance(mesh_or_vertices, trimesh.Trimesh): pyrender_mesh = self._create_mesh_from_trimesh(mesh_or_vertices, color) else: # Assume it's a vertices array pyrender_mesh = self._create_mesh_from_vertices(mesh_or_vertices, color) mesh_node = self.scene.add(pyrender_mesh) cam = WeakPerspectiveCamera(scale=camera_params[:2], translation=camera_params[2:]) cam_node = self.scene.add(cam, pose=np.eye(4)) if depth_only: depth = self.renderer.render(self.scene, flags=pyrender.RenderFlags.DEPTH_ONLY) self.scene.remove_node(mesh_node); self.scene.remove_node(cam_node) return depth img_rgba, depth = self.renderer.render(self.scene, flags=pyrender.RenderFlags.RGBA) # Make a writable copy since pyrender returns read-only arrays img_rgba = img_rgba.copy() alpha = (depth > 0).astype(np.uint8) * 255 img_rgba[..., 3] = alpha self.scene.remove_node(mesh_node); self.scene.remove_node(cam_node) return img_rgba def overlay_rgba_on_rgb(rendered_rgba, background_rgb): fg = Image.fromarray(rendered_rgba, mode="RGBA") bg = Image.fromarray(background_rgb, mode="RGB") if fg.size != bg.size: bg = bg.resize(fg.size, resample=Image.BILINEAR) out = Image.alpha_composite(bg.convert("RGBA"), fg) return np.array(out.convert("RGB"))