Utils.py

# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto.  Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.


import os, sys, time,torch,pickle,trimesh,itertools,pdb,zipfile,datetime,imageio,gzip,logging,joblib,importlib,uuid,signal,multiprocessing,psutil,subprocess,tarfile,scipy,argparse
from pytorch3d.transforms import so3_log_map,so3_exp_map,se3_exp_map,se3_log_map,matrix_to_axis_angle,matrix_to_euler_angles,euler_angles_to_matrix, rotation_6d_to_matrix
from pytorch3d.renderer import FoVPerspectiveCameras, PerspectiveCameras, look_at_view_transform, look_at_rotation, RasterizationSettings, MeshRenderer, MeshRasterizer, BlendParams, SoftSilhouetteShader, HardPhongShader, PointLights, TexturesVertex
from pytorch3d.renderer.mesh.rasterize_meshes import barycentric_coordinates
from pytorch3d.renderer.mesh.shader import SoftDepthShader, HardFlatShader
from pytorch3d.renderer.mesh.textures import Textures
from pytorch3d.structures import Meshes
from scipy.interpolate import griddata
import nvdiffrast.torch as dr
import torch.nn.functional as F
import torchvision
import torch.nn as nn
from functools import partial
import pandas as pd
import open3d as o3d
from uuid import uuid4
import cv2
from PIL import Image
import numpy as np
from collections import defaultdict
import multiprocessing as mp
import matplotlib.pyplot as plt
import math,glob,re,copy
from transformations import *
from scipy.spatial import cKDTree
from collections import OrderedDict
import ruamel.yaml
yaml = ruamel.yaml.YAML()
code_dir = os.path.dirname(os.path.realpath(__file__))
sys.path.append(code_dir)
# sys.path.append(f"{code_dir}/mycpp/build")
try:
  import kornia
except:
  kornia = None
try:
  import mycpp.build.mycpp as mycpp
except:
  mycpp = None
try:
  from bundlesdf.mycuda import common
except:
  common = None
try:
  import warp as wp
  wp.init()
except:
  wp = None
enable_timer = 0

def NestDict():
  return defaultdict(NestDict)

to8b = lambda x : (255*np.clip(x,0,1)).astype(np.uint8)

BAD_DEPTH = 99
BAD_COLOR = 0

glcam_in_cvcam = np.array([[1,0,0,0],
                          [0,-1,0,0],
                          [0,0,-1,0],
                          [0,0,0,1]]).astype(float)

COLOR_MAP=np.array([[0, 0, 0], #Ignore
                    [128,0,0], #Background
                    [0,128,0], #Wall
                    [128,128,0], #Floor
                    [0,0,128], #Ceiling
                    [128,0,128], #Table
                    [0,128,128], #Chair
                    [128,128,128], #Window
                    [64,0,0], #Door
                    [192,0,0], #Monitor
                    [64, 128, 0],     # 11th
                    [192, 0, 128],
                    [64, 128, 128],
                    [192, 128, 128],
                    [0, 64, 0],
                    [128, 64, 0],
                    [0, 192, 0],
                    [128, 192, 0],
                    ])


def set_logging_format(level=logging.INFO):
  importlib.reload(logging)
  FORMAT = '[%(funcName)s()] %(message)s'
  logging.basicConfig(level=level, format=FORMAT)

set_logging_format()




def make_mesh_tensors(mesh, device='cuda', max_tex_size=None):
  mesh_tensors = {}
  if isinstance(mesh.visual, trimesh.visual.texture.TextureVisuals):
    img = np.array(mesh.visual.material.image.convert('RGB'))
    img = img[...,:3]
    if max_tex_size is not None:
      max_size = max(img.shape[0], img.shape[1])
      if max_size>max_tex_size:
        scale = 1/max_size * max_tex_size
        img = cv2.resize(img, fx=scale, fy=scale, dsize=None)
    mesh_tensors['tex'] = torch.as_tensor(img, device=device, dtype=torch.float)[None]/255.0
    mesh_tensors['uv_idx']  = torch.as_tensor(mesh.faces, device=device, dtype=torch.int)
    uv = torch.as_tensor(mesh.visual.uv, device=device, dtype=torch.float)
    uv[:,1] = 1 - uv[:,1]
    mesh_tensors['uv']  = uv
  else:
    if mesh.visual.vertex_colors is None:
      logging.info(f"WARN: mesh doesn't have vertex_colors, assigning a pure color")
      mesh.visual.vertex_colors = np.tile(np.array([128,128,128]).reshape(1,3), (len(mesh.vertices), 1))
    mesh_tensors['vertex_color'] = torch.as_tensor(mesh.visual.vertex_colors[...,:3], device=device, dtype=torch.float)/255.0

  mesh_tensors.update({
    'pos': torch.tensor(mesh.vertices, device=device, dtype=torch.float),
    'faces': torch.tensor(mesh.faces, device=device, dtype=torch.int),
    'vnormals': torch.tensor(mesh.vertex_normals, device=device, dtype=torch.float),
  })
  return mesh_tensors


def nvdiffrast_render(K=None, H=None, W=None, ob_in_cams=None, glctx=None, context='cuda', get_normal=False, mesh_tensors=None, mesh=None, projection_mat=None, bbox2d=None, output_size=None, use_light=False, light_color=None, light_dir=np.array([0,0,1]), light_pos=np.array([0,0,0]), w_ambient=0.8, w_diffuse=0.5, extra={}):
  '''Just plain rendering, not support any gradient
  @K: (3,3) np array
  @ob_in_cams: (N,4,4) torch tensor, openCV camera
  @projection_mat: np array (4,4)
  @output_size: (height, width)
  @bbox2d: (N,4) (umin,vmin,umax,vmax) if only roi need to render.
  @light_dir: in cam space
  @light_pos: in cam space
  '''
  if glctx is None:
    if context == 'gl':
      glctx = dr.RasterizeGLContext()
    elif context=='cuda':
      glctx = dr.RasterizeCudaContext()
    else:
      raise NotImplementedError
    logging.info("created context")

  if mesh_tensors is None:
    mesh_tensors = make_mesh_tensors(mesh)
  pos = mesh_tensors['pos']
  vnormals = mesh_tensors['vnormals']
  pos_idx = mesh_tensors['faces']
  has_tex = 'tex' in mesh_tensors

  ob_in_glcams = torch.tensor(glcam_in_cvcam, device='cuda', dtype=torch.float)[None]@ob_in_cams
  if projection_mat is None:
    projection_mat = projection_matrix_from_intrinsics(K, height=H, width=W, znear=0.1, zfar=100)
  projection_mat = torch.as_tensor(projection_mat.reshape(-1,4,4), device='cuda', dtype=torch.float)
  mtx = projection_mat@ob_in_glcams

  if output_size is None:
    output_size = np.asarray([H,W])

  pts_cam = transform_pts(pos, ob_in_cams)
  pos_homo = to_homo_torch(pos)
  pos_clip = (mtx[:,None]@pos_homo[None,...,None])[...,0]
  if bbox2d is not None:
    l = bbox2d[:,0]
    t = H-bbox2d[:,1]
    r = bbox2d[:,2]
    b = H-bbox2d[:,3]
    tf = torch.eye(4, dtype=torch.float, device='cuda').reshape(1,4,4).expand(len(ob_in_cams),4,4).contiguous()
    tf[:,0,0] = W/(r-l)
    tf[:,1,1] = H/(t-b)
    tf[:,3,0] = (W-r-l)/(r-l)
    tf[:,3,1] = (H-t-b)/(t-b)
    pos_clip = pos_clip@tf
  rast_out, _ = dr.rasterize(glctx, pos_clip, pos_idx, resolution=np.asarray(output_size))
  xyz_map, _ = dr.interpolate(pts_cam, rast_out, pos_idx)
  depth = xyz_map[...,2]
  if has_tex:
    texc, _ = dr.interpolate(mesh_tensors['uv'], rast_out, mesh_tensors['uv_idx'])
    color = dr.texture(mesh_tensors['tex'], texc, filter_mode='linear')
  else:
    color, _ = dr.interpolate(mesh_tensors['vertex_color'], rast_out, pos_idx)

  if use_light:
    get_normal = True
  if get_normal:
    vnormals_cam = transform_dirs(vnormals, ob_in_cams)
    normal_map, _ = dr.interpolate(vnormals_cam, rast_out, pos_idx)
    normal_map = F.normalize(normal_map, dim=-1)
    normal_map = torch.flip(normal_map, dims=[1])
  else:
    normal_map = None

  if use_light:
    if light_dir is not None:
      light_dir_neg = -torch.as_tensor(light_dir, dtype=torch.float, device='cuda')
    else:
      light_dir_neg = torch.as_tensor(light_pos, dtype=torch.float, device='cuda').reshape(1,1,3) - pts_cam
    diffuse_intensity = (F.normalize(vnormals_cam, dim=-1) * F.normalize(light_dir_neg, dim=-1)).sum(dim=-1).clip(0, 1)[...,None]
    diffuse_intensity_map, _ = dr.interpolate(diffuse_intensity, rast_out, pos_idx)  # (N_pose, H, W, 1)
    if light_color is None:
      light_color = color
    else:
      light_color = torch.as_tensor(light_color, device='cuda', dtype=torch.float)
    color = color*w_ambient + diffuse_intensity_map*light_color*w_diffuse

  color = color.clip(0,1)
  color = color * torch.clamp(rast_out[..., -1:], 0, 1) # Mask out background using alpha
  color = torch.flip(color, dims=[1])   # Flip Y coordinates
  depth = torch.flip(depth, dims=[1])
  extra['xyz_map'] = torch.flip(xyz_map, dims=[1])
  return color, depth, normal_map


def set_seed(random_seed):
  import torch,random
  np.random.seed(random_seed)
  random.seed(random_seed)
  torch.manual_seed(random_seed)
  torch.cuda.manual_seed_all(random_seed)
  torch.backends.cudnn.deterministic = True
  torch.backends.cudnn.benchmark = False


def add_err(pred,gt,model_pts,symetry_tfs=np.eye(4)[None]):
  """
  Average Distance of Model Points for objects with no indistinguishable views
  - by Hinterstoisser et al. (ACCV 2012).
  """
  pred_pts = transform_pts(model_pts, pred)
  gt_pts = transform_pts(model_pts, gt)
  e = np.linalg.norm(pred_pts - gt_pts, axis=-1).mean()
  return e

def adds_err(pred,gt,model_pts):
  """
  @pred: 4x4 mat
  @gt:
  @model: (N,3)
  """
  pred_pts = transform_pts(model_pts, pred)
  gt_pts = transform_pts(model_pts, gt)
  nn_index = cKDTree(pred_pts)
  nn_dists, _ = nn_index.query(gt_pts, k=1, workers=-1)
  e = nn_dists.mean()
  return e

def compute_auc_sklearn(errs, max_val=0.1, step=0.001):
  from sklearn import metrics
  errs = np.sort(np.array(errs))
  X = np.arange(0, max_val+step, step)
  Y = np.ones(len(X))
  for i,x in enumerate(X):
    y = (errs<=x).sum()/len(errs)
    Y[i] = y
    if y>=1:
      break
  auc = metrics.auc(X, Y) / (max_val*1)
  return auc



def normalizeRotation(pose):
  '''Assume no shear case
  '''
  new_pose = pose.copy()
  scales = np.linalg.norm(pose[:3,:3],axis=0)
  new_pose[:3,:3] /= scales.reshape(1,3)
  return new_pose



def toOpen3dCloud(points,colors=None,normals=None):
  cloud = o3d.geometry.PointCloud()
  cloud.points = o3d.utility.Vector3dVector(points.astype(np.float64))
  if colors is not None:
    if colors.max()>1:
      colors = colors/255.0
    cloud.colors = o3d.utility.Vector3dVector(colors.astype(np.float64))
  if normals is not None:
    cloud.normals = o3d.utility.Vector3dVector(normals.astype(np.float64))
  return cloud



def make_grid_image(imgs, nrow, padding=5, pad_value=255):
  '''
  @imgs: (B,H,W,C) np array
  @nrow: num of images per row
  '''
  grid = torchvision.utils.make_grid(torch.as_tensor(np.asarray(imgs)).permute(0,3,1,2), nrow=nrow, padding=padding, pad_value=pad_value)
  grid = grid.permute(1,2,0).contiguous().data.cpu().numpy().astype(np.uint8)
  return grid


if wp is not None:
  @wp.kernel(enable_backward=False)
  def bilateral_filter_depth_kernel(depth:wp.array(dtype=float, ndim=2), out:wp.array(dtype=float, ndim=2), radius:int, zfar:float, sigmaD:float, sigmaR:float):
    h,w = wp.tid()
    H = depth.shape[0]
    W = depth.shape[1]
    if w>=W or h>=H:
      return
    out[h,w] = 0.0
    mean_depth = float(0.0)
    num_valid = int(0)
    for u in range(w-radius, w+radius+1):
      if u<0 or u>=W:
        continue
      for v in range(h-radius, h+radius+1):
        if v<0 or v>=H:
          continue
        cur_depth = depth[v,u]
        if cur_depth>=0.1 and cur_depth<zfar:
          num_valid += 1
          mean_depth += cur_depth
    if num_valid==0:
      return
    mean_depth /= float(num_valid)

    depthCenter = depth[h,w]
    sum_weight = float(0.0)
    sum = float(0.0)
    for u in range(w-radius, w+radius+1):
      if u<0 or u>=W:
        continue
      for v in range(h-radius, h+radius+1):
        if v<0 or v>=H:
          continue
        cur_depth = depth[v,u]
        if cur_depth>=0.1 and cur_depth<zfar and abs(cur_depth-mean_depth)<0.01:
          weight = wp.exp( -float((u-w)*(u-w) + (h-v)*(h-v)) / (2.0*sigmaD*sigmaD) - (depthCenter-cur_depth)*(depthCenter-cur_depth)/(2.0*sigmaR*sigmaR) )
          sum_weight += weight
          sum += weight*cur_depth
    if sum_weight>0 and num_valid>0:
      out[h,w] = sum/sum_weight

  def bilateral_filter_depth(depth, radius=2, zfar=100, sigmaD=2, sigmaR=100000, device='cuda'):
    if isinstance(depth, np.ndarray):
      depth_wp = wp.array(depth, dtype=float, device=device)
    else:
      depth_wp = wp.from_torch(depth)
    out_wp = wp.zeros(depth.shape, dtype=float, device=device)
    wp.launch(kernel=bilateral_filter_depth_kernel, device=device, dim=[depth.shape[0], depth.shape[1]], inputs=[depth_wp, out_wp, radius, zfar, sigmaD, sigmaR])
    depth_out = wp.to_torch(out_wp)

    if isinstance(depth, np.ndarray):
      depth_out = depth_out.data.cpu().numpy()
    return depth_out


  @wp.kernel(enable_backward=False)
  def erode_depth_kernel(depth:wp.array(dtype=float, ndim=2), out:wp.array(dtype=float, ndim=2), radius:int, depth_diff_thres:float, ratio_thres:float, zfar:float):
    h,w = wp.tid()
    H = depth.shape[0]
    W = depth.shape[1]
    if w>=W or h>=H:
      return
    d_ori = depth[h,w]
    if d_ori<0.1 or d_ori>=zfar:
      out[h,w] = 0.0
    bad_cnt = float(0)
    total = float(0)
    for u in range(w-radius, w+radius+1):
      if u<0 or u>=W:
        continue
      for v in range(h-radius, h+radius+1):
        if v<0 or v>=H:
          continue
        cur_depth = depth[v,u]
        total += 1.0
        if cur_depth<0.1 or cur_depth>=zfar or abs(cur_depth-d_ori)>depth_diff_thres:
          bad_cnt += 1.0
    if bad_cnt/total>ratio_thres:
      out[h,w] = 0.0
    else:
      out[h,w] = d_ori


  def erode_depth(depth, radius=2, depth_diff_thres=0.001, ratio_thres=0.8, zfar=100, device='cuda'):
    depth_wp = wp.from_torch(torch.as_tensor(depth, dtype=torch.float, device=device))
    out_wp = wp.zeros(depth.shape, dtype=float, device=device)
    wp.launch(kernel=erode_depth_kernel, device=device, dim=[depth.shape[0], depth.shape[1]], inputs=[depth_wp, out_wp, radius, depth_diff_thres, ratio_thres, zfar],)
    depth_out = wp.to_torch(out_wp)

    if isinstance(depth, np.ndarray):
      depth_out = depth_out.data.cpu().numpy()
    return depth_out



def depth2xyzmap(depth, K, uvs=None):
  invalid_mask = (depth<0.1)
  H,W = depth.shape[:2]
  if uvs is None:
    vs,us = np.meshgrid(np.arange(0,H),np.arange(0,W), sparse=False, indexing='ij')
    vs = vs.reshape(-1)
    us = us.reshape(-1)
  else:
    uvs = uvs.round().astype(int)
    us = uvs[:,0]
    vs = uvs[:,1]
  zs = depth[vs,us]
  xs = (us-K[0,2])*zs/K[0,0]
  ys = (vs-K[1,2])*zs/K[1,1]
  pts = np.stack((xs.reshape(-1),ys.reshape(-1),zs.reshape(-1)), 1)  #(N,3)
  xyz_map = np.zeros((H,W,3), dtype=np.float32)
  xyz_map[vs,us] = pts
  xyz_map[invalid_mask] = 0
  return xyz_map


def depth2xyzmap_batch(depths, Ks, zfar):
  '''
  @depths: torch tensor (B,H,W)
  @Ks: torch tensor (B,3,3)
  '''
  bs = depths.shape[0]
  invalid_mask = (depths<0.1) | (depths>zfar)
  H,W = depths.shape[-2:]
  vs,us = torch.meshgrid(torch.arange(0,H),torch.arange(0,W), indexing='ij')
  vs = vs.reshape(-1).float().cuda()[None].expand(bs,-1)
  us = us.reshape(-1).float().cuda()[None].expand(bs,-1)
  zs = depths.reshape(bs,-1)
  Ks = Ks[:,None].expand(bs,zs.shape[-1],3,3)
  xs = (us-Ks[...,0,2])*zs/Ks[...,0,0]  #(B,N)
  ys = (vs-Ks[...,1,2])*zs/Ks[...,1,1]
  pts = torch.stack([xs,ys,zs], dim=-1)  #(B,N,3)
  xyz_maps = pts.reshape(bs,H,W,3)
  xyz_maps[invalid_mask] = 0
  return xyz_maps



def rle_to_mask(rle: dict) -> np.ndarray:
  """Compute a binary mask from an uncompressed RLE."""
  h, w = rle["size"]
  mask = np.empty(h * w, dtype=bool)
  idx = 0
  parity = False
  for count in rle["counts"]:
      mask[idx : idx + count] = parity
      idx += count
      parity ^= True
  mask = mask.reshape(w, h)
  return mask.transpose()  # Put in C order


def depth_to_vis(depth, zmin=None, zmax=None, mode='rgb', inverse=True):
  if zmin is None:
    zmin = depth.min()
  if zmax is None:
    zmax = depth.max()

  if inverse:
    invalid = depth<0.1
    vis = zmin/(depth+1e-8)
    vis[invalid] = 0
  else:
    depth = depth.clip(zmin, zmax)
    invalid = (depth==zmin) | (depth==zmax)
    vis = (depth-zmin)/(zmax-zmin)
    vis[invalid] = 1

  if mode=='gray':
    vis = (vis*255).clip(0, 255).astype(np.uint8)
  elif mode=='rgb':
    vis = cv2.applyColorMap((vis*255).astype(np.uint8), cv2.COLORMAP_JET)[...,::-1]
  else:
    raise RuntimeError

  return vis



def sample_views_icosphere(n_views, subdivisions=None, radius=1):
  if subdivisions is not None:
    mesh = trimesh.creation.icosphere(subdivisions=subdivisions, radius=radius)
  else:
    subdivision = 1
    while 1:
      mesh = trimesh.creation.icosphere(subdivisions=subdivision, radius=radius)
      if mesh.vertices.shape[0]>=n_views:
        break
      subdivision += 1
  cam_in_obs = np.tile(np.eye(4)[None], (len(mesh.vertices),1,1))
  cam_in_obs[:,:3,3] = mesh.vertices
  up = np.array([0,0,1])
  z_axis = -cam_in_obs[:,:3,3]  #(N,3)
  z_axis /= np.linalg.norm(z_axis, axis=-1).reshape(-1,1)
  x_axis = np.cross(up.reshape(1,3), z_axis)
  invalid = (x_axis==0).all(axis=-1)
  x_axis[invalid] = [1,0,0]
  x_axis /= np.linalg.norm(x_axis, axis=-1).reshape(-1,1)
  y_axis = np.cross(z_axis, x_axis)
  y_axis /= np.linalg.norm(y_axis, axis=-1).reshape(-1,1)
  cam_in_obs[:,:3,0] = x_axis
  cam_in_obs[:,:3,1] = y_axis
  cam_in_obs[:,:3,2] = z_axis
  return cam_in_obs



def to_homo(pts):
  '''
  @pts: (N,3 or 2) will homogeneliaze the last dimension
  '''
  assert len(pts.shape)==2, f'pts.shape: {pts.shape}'
  homo = np.concatenate((pts, np.ones((pts.shape[0],1))),axis=-1)
  return homo


def to_homo_torch(pts):
  '''
  @pts: shape can be (...,N,3 or 2) or (N,3) will homogeneliaze the last dimension
  '''
  ones = torch.ones((*pts.shape[:-1],1), dtype=torch.float, device=pts.device)
  homo = torch.cat((pts, ones),dim=-1)
  return homo


def transform_pts(pts,tf):
  """Transform 2d or 3d points
  @pts: (...,N_pts,3)
  @tf: (...,4,4)
  """
  if len(tf.shape)>=3 and tf.shape[-3]!=pts.shape[-2]:
    tf = tf[...,None,:,:]
  return (tf[...,:-1,:-1]@pts[...,None] + tf[...,:-1,-1:])[...,0]


def transform_dirs(dirs,tf):
  """
  @dirs: (...,3)
  @tf: (...,4,4)
  """
  if len(tf.shape)>=3 and tf.shape[-3]!=dirs.shape[-2]:
    tf = tf[...,None,:,:]
  return (tf[...,:3,:3]@dirs[...,None])[...,0]



def random_direction():
  '''https://stackoverflow.com/questions/33976911/generate-a-random-sample-of-points-distributed-on-the-surface-of-a-unit-sphere
  '''
  vec = np.random.randn(3).reshape(3)
  vec /= np.linalg.norm(vec)
  return vec



def compute_mesh_diameter(model_pts=None, mesh=None, n_sample=1000):
  from sklearn.decomposition import TruncatedSVD
  if mesh is not None:
    u, s, vh = scipy.linalg.svd(mesh.vertices, full_matrices=False)
    pts = u@s
    diameter = np.linalg.norm(pts.max(axis=0)-pts.min(axis=0))
    return float(diameter)

  if n_sample is None:
    pts = model_pts
  else:
    ids = np.random.choice(len(model_pts), size=min(n_sample, len(model_pts)), replace=False)
    pts = model_pts[ids]
  dists = np.linalg.norm(pts[None]-pts[:,None], axis=-1)
  diameter = dists.max()
  return diameter


def compute_crop_window_tf_batch(pts=None, H=None, W=None, poses=None, K=None, crop_ratio=1.2, out_size=None, rgb=None, uvs=None, method='min_box', mesh_diameter=None):
  '''Project the points and find the cropping transform
  @pts: (N,3)
  @poses: (B,4,4) tensor
  @min_box: min_box/min_circle
  @scale: scale to apply to the tightly enclosing roi
  '''
  def compute_tf_batch(left, right, top, bottom):
    B = len(left)
    left = left.round()
    right = right.round()
    top = top.round()
    bottom = bottom.round()

    tf = torch.eye(3)[None].expand(B,-1,-1).contiguous()
    tf[:,0,2] = -left
    tf[:,1,2] = -top
    new_tf = torch.eye(3)[None].expand(B,-1,-1).contiguous()
    new_tf[:,0,0] = out_size[0]/(right-left)
    new_tf[:,1,1] = out_size[1]/(bottom-top)
    tf = new_tf@tf
    return tf

  B = len(poses)
  torch.set_default_tensor_type('torch.cuda.FloatTensor')
  if method=='box_3d':
    radius = mesh_diameter*crop_ratio/2
    offsets = torch.tensor([0,0,0,
                        radius,0,0,
                        -radius,0,0,
                        0,radius,0,
                        0,-radius,0]).reshape(-1,3)
    pts = poses[:,:3,3].reshape(-1,1,3)+offsets.reshape(1,-1,3)
    K = torch.as_tensor(K)
    projected = (K@pts.reshape(-1,3).T).T
    uvs = projected[:,:2]/projected[:,2:3]
    uvs = uvs.reshape(B, -1, 2)
    center = uvs[:,0]  #(B,2)
    radius = torch.abs(uvs-center.reshape(-1,1,2)).reshape(B,-1).max(axis=-1)[0].reshape(-1)  #(B)
    left = center[:,0]-radius
    right = center[:,0]+radius
    top = center[:,1]-radius
    bottom = center[:,1]+radius
    tfs = compute_tf_batch(left, right, top, bottom)
    return tfs

  else:
    raise RuntimeError

  return tf



def cv_draw_text(img,text,uv_top_left,color=(255, 255, 255),fontScale=0.5,thickness=1,fontFace=cv2.FONT_HERSHEY_SIMPLEX,outline_color=None,line_spacing=1.5):
  H,W = img.shape[:2]
  uv_top_left = np.array(uv_top_left, dtype=float)
  assert uv_top_left.shape == (2,)

  for line in text.splitlines():
    (w, h), _ = cv2.getTextSize(text=line,fontFace=fontFace,fontScale=fontScale,thickness=thickness,)
    uv_bottom_left_i = uv_top_left + [0, h]

    ############# Ensure inside image
    while uv_bottom_left_i[0]<0:
      uv_bottom_left_i[0] += 1
    while uv_bottom_left_i[0]+w>=W:
      uv_bottom_left_i[0] -= 1
    while uv_bottom_left_i[1]>=H:
      uv_bottom_left_i[1] -= 1
    while uv_bottom_left_i[1]-h<0:
      uv_bottom_left_i[1] += 1

    org = tuple(uv_bottom_left_i.astype(int))

    if outline_color is not None:
      cv2.putText(img,text=line,org=org,fontFace=fontFace,fontScale=fontScale,color=outline_color,thickness=thickness,lineType=cv2.LINE_AA,)
    cv2.putText(img,text=line,org=org,fontFace=fontFace,fontScale=fontScale,color=color,thickness=thickness,lineType=cv2.LINE_AA,)
    uv_top_left[1] =  uv_bottom_left_i[1]-h+h*line_spacing
  return img


def trimesh_add_pure_colored_texture(mesh, color=np.array([255,255,255]), resolution=5):
  tex_img = np.tile(color.reshape(1,1,3), (resolution, resolution, 1)).astype(np.uint8)
  mesh = mesh.unwrap()
  mesh.visual = trimesh.visual.texture.TextureVisuals(uv=mesh.visual.uv,image=Image.fromarray(tex_img))
  return mesh




def project_3d_to_2d(pt,K,ob_in_cam):
  pt = pt.reshape(4,1)
  projected = K @ ((ob_in_cam@pt)[:3,:])
  projected = projected.reshape(-1)
  projected = projected/projected[2]
  return projected.reshape(-1)[:2].round().astype(int)


def draw_xyz_axis(color, ob_in_cam, scale=0.1, K=np.eye(3), thickness=3, transparency=0,is_input_rgb=False):
  '''
  @color: BGR
  '''
  if is_input_rgb:
    color = cv2.cvtColor(color,cv2.COLOR_RGB2BGR)
  xx = np.array([1,0,0,1]).astype(float)
  yy = np.array([0,1,0,1]).astype(float)
  zz = np.array([0,0,1,1]).astype(float)
  xx[:3] = xx[:3]*scale
  yy[:3] = yy[:3]*scale
  zz[:3] = zz[:3]*scale
  origin = tuple(project_3d_to_2d(np.array([0,0,0,1]), K, ob_in_cam))
  xx = tuple(project_3d_to_2d(xx, K, ob_in_cam))
  yy = tuple(project_3d_to_2d(yy, K, ob_in_cam))
  zz = tuple(project_3d_to_2d(zz, K, ob_in_cam))
  line_type = cv2.LINE_AA
  arrow_len = 0
  tmp = color.copy()
  tmp1 = tmp.copy()
  tmp1 = cv2.arrowedLine(tmp1, origin, xx, color=(0,0,255), thickness=thickness,line_type=line_type, tipLength=arrow_len)
  mask = np.linalg.norm(tmp1-tmp, axis=-1)>0
  tmp[mask] = tmp[mask]*transparency + tmp1[mask]*(1-transparency)
  tmp1 = tmp.copy()
  tmp1 = cv2.arrowedLine(tmp1, origin, yy, color=(0,255,0), thickness=thickness,line_type=line_type, tipLength=arrow_len)
  mask = np.linalg.norm(tmp1-tmp, axis=-1)>0
  tmp[mask] = tmp[mask]*transparency + tmp1[mask]*(1-transparency)
  tmp1 = tmp.copy()
  tmp1 = cv2.arrowedLine(tmp1, origin, zz, color=(255,0,0), thickness=thickness,line_type=line_type, tipLength=arrow_len)
  mask = np.linalg.norm(tmp1-tmp, axis=-1)>0
  tmp[mask] = tmp[mask]*transparency + tmp1[mask]*(1-transparency)
  tmp = tmp.astype(np.uint8)
  if is_input_rgb:
    tmp = cv2.cvtColor(tmp,cv2.COLOR_BGR2RGB)

  return tmp


def draw_posed_3d_box(K, img, ob_in_cam, bbox, line_color=(0,255,0), linewidth=2):
  '''Revised from 6pack dataset/inference_dataset_nocs.py::projection
  @bbox: (2,3) min/max
  @line_color: RGB
  '''
  min_xyz = bbox.min(axis=0)
  xmin, ymin, zmin = min_xyz
  max_xyz = bbox.max(axis=0)
  xmax, ymax, zmax = max_xyz

  def draw_line3d(start,end,img):
    pts = np.stack((start,end),axis=0).reshape(-1,3)
    pts = (ob_in_cam@to_homo(pts).T).T[:,:3]   #(2,3)
    projected = (K@pts.T).T
    uv = np.round(projected[:,:2]/projected[:,2].reshape(-1,1)).astype(int)   #(2,2)
    img = cv2.line(img, uv[0].tolist(), uv[1].tolist(), color=line_color, thickness=linewidth, lineType=cv2.LINE_AA)
    return img

  for y in [ymin,ymax]:
    for z in [zmin,zmax]:
      start = np.array([xmin,y,z])
      end = start+np.array([xmax-xmin,0,0])
      img = draw_line3d(start,end,img)

  for x in [xmin,xmax]:
    for z in [zmin,zmax]:
      start = np.array([x,ymin,z])
      end = start+np.array([0,ymax-ymin,0])
      img = draw_line3d(start,end,img)

  for x in [xmin,xmax]:
    for y in [ymin,ymax]:
      start = np.array([x,y,zmin])
      end = start+np.array([0,0,zmax-zmin])
      img = draw_line3d(start,end,img)

  return img


def projection_matrix_from_intrinsics(K, height, width, znear, zfar, window_coords='y_down'):
  """Conversion of Hartley-Zisserman intrinsic matrix to OpenGL proj. matrix.

  Ref:
  1) https://strawlab.org/2011/11/05/augmented-reality-with-OpenGL
  2) https://github.com/strawlab/opengl-hz/blob/master/src/calib_test_utils.py

  :param K: 3x3 ndarray with the intrinsic camera matrix.
  :param x0 The X coordinate of the camera image origin (typically 0).
  :param y0: The Y coordinate of the camera image origin (typically 0).
  :param w: Image width.
  :param h: Image height.
  :param nc: Near clipping plane.
  :param fc: Far clipping plane.
  :param window_coords: 'y_up' or 'y_down'.
  :return: 4x4 ndarray with the OpenGL projection matrix.
  """
  x0 = 0
  y0 = 0
  w = width
  h = height
  nc = znear
  fc = zfar

  depth = float(fc - nc)
  q = -(fc + nc) / depth
  qn = -2 * (fc * nc) / depth

  # Draw our images upside down, so that all the pixel-based coordinate
  # systems are the same.
  if window_coords == 'y_up':
    proj = np.array([
      [2 * K[0, 0] / w, -2 * K[0, 1] / w, (-2 * K[0, 2] + w + 2 * x0) / w, 0],
      [0, -2 * K[1, 1] / h, (-2 * K[1, 2] + h + 2 * y0) / h, 0],
      [0, 0, q, qn],  # Sets near and far planes (glPerspective).
      [0, 0, -1, 0]
      ])

  # Draw the images upright and modify the projection matrix so that OpenGL
  # will generate window coords that compensate for the flipped image coords.
  elif window_coords == 'y_down':
    proj = np.array([
      [2 * K[0, 0] / w, -2 * K[0, 1] / w, (-2 * K[0, 2] + w + 2 * x0) / w, 0],
      [0, 2 * K[1, 1] / h, (2 * K[1, 2] - h + 2 * y0) / h, 0],
      [0, 0, q, qn],  # Sets near and far planes (glPerspective).
      [0, 0, -1, 0]
      ])
  else:
    raise NotImplementedError

  return proj



def symmetry_tfs_from_info(info, rot_angle_discrete=5):
  symmetry_tfs = [np.eye(4)]
  if 'symmetries_discrete' in info:
    tfs = np.array(info['symmetries_discrete']).reshape(-1,4,4)
    tfs[...,:3,3] *= 0.001
    symmetry_tfs = [np.eye(4)]
    symmetry_tfs += list(tfs)
  if 'symmetries_continuous' in info:
    axis = np.array(info['symmetries_continuous'][0]['axis']).reshape(3)
    offset = info['symmetries_continuous'][0]['offset']
    rxs = [0]
    rys = [0]
    rzs = [0]
    if axis[0]>0:
      rxs = np.arange(0,360,rot_angle_discrete)/180.0*np.pi
    elif axis[1]>0:
      rys = np.arange(0,360,rot_angle_discrete)/180.0*np.pi
    elif axis[2]>0:
      rzs = np.arange(0,360,rot_angle_discrete)/180.0*np.pi
    for rx in rxs:
      for ry in rys:
        for rz in rzs:
          tf = euler_matrix(rx, ry, rz)
          tf[:3,3] = offset
          symmetry_tfs.append(tf)
  if len(symmetry_tfs)==0:
    symmetry_tfs = [np.eye(4)]
  symmetry_tfs = np.array(symmetry_tfs)
  return symmetry_tfs



def pose_to_egocentric_delta_pose(A_in_cam, B_in_cam):
  '''Used for Pose Refinement. Given the object's two poses in camera, convert them to relative poses in camera's egocentric view
  @A_in_cam: (B,4,4) torch tensor
  '''
  trans_delta = B_in_cam[:,:3,3] - A_in_cam[:,:3,3]
  rot_mat_delta = B_in_cam[:,:3,:3]@A_in_cam[:,:3,:3].permute(0,2,1)
  return trans_delta, rot_mat_delta



def egocentric_delta_pose_to_pose(A_in_cam, trans_delta, rot_mat_delta):
  '''Used for Pose Refinement. Given the object's two poses in camera, convert them to relative poses in camera's egocentric view
  @A_in_cam: (B,4,4) torch tensor
  '''
  B_in_cam = torch.eye(4, dtype=torch.float, device=A_in_cam.device)[None].expand(len(A_in_cam),-1,-1).contiguous()
  B_in_cam[:,:3,3] = A_in_cam[:,:3,3]+trans_delta
  B_in_cam[:,:3,:3] = rot_mat_delta@A_in_cam[:,:3,:3]
  return B_in_cam


def sdg_load_bounding_box(file_path: str):
  """Load bounding boxes.
  Args:
      file_path: Path of the bounding box.

  Returns:
      A dictionary of the bounding boxes.
  """
  bbox_dict = {}
  bbox_array = np.load(file_path)
  for id, x_min, y_min, x_max, y_max, occlusion_ratio in zip(
      bbox_array["semanticId"],
      bbox_array["x_min"],
      bbox_array["y_min"],
      bbox_array["x_max"],
      bbox_array["y_max"],
      bbox_array["occlusionRatio"],
  ):
      bbox_dict[id] = {
          "x_min": x_min,
          "y_min": y_min,
          "x_max": x_max,
          "y_max": y_max,
          "occlusion_ratio": occlusion_ratio,
      }
  return bbox_dict


def texture_map_interpolation(tex_image_numpy):
  all_channels = []
  mask = np.all(tex_image_numpy == 0, axis=2)
  x = np.arange(0, tex_image_numpy.shape[1])
  y = np.arange(0, tex_image_numpy.shape[0])
  xx, yy = np.meshgrid(x, y)
  for each_channel in range(tex_image_numpy.shape[2]):
      curr_channel = tex_image_numpy[:,:,each_channel]
      x1 = xx[~mask]
      y1 = yy[~mask]
      newarr = curr_channel[~mask]
      GD1 = griddata((x1, y1), newarr.ravel(), (xx, yy), method='nearest')
      all_channels.append(GD1[:,:,np.newaxis].round().astype(np.uint8))
  final_image = np.concatenate(all_channels, axis =-1)
  return final_image



class OctreeManager:
  def __init__(self,pts=None,max_level=None,octree=None):
    import kaolin
    if octree is None:
      pts_quantized = kaolin.ops.spc.quantize_points(pts.contiguous(), level=max_level)
      self.octree = kaolin.ops.spc.unbatched_points_to_octree(pts_quantized, max_level, sorted=False)
    else:
      self.octree = octree
    lengths = torch.tensor([len(self.octree)], dtype=torch.int32).cpu()
    self.max_level, self.pyramids, self.exsum = kaolin.ops.spc.scan_octrees(self.octree,lengths)
    self.finest_vox_size = 2.0/(2**self.max_level)
    self.n_level = self.max_level+1
    self.vox_point_all_levels = kaolin.ops.spc.generate_points(self.octree, self.pyramids, self.exsum)
    self.point_hierarchy_dual, self.pyramid_dual = kaolin.ops.spc.unbatched_make_dual(self.vox_point_all_levels, self.pyramids[0])
    self.trinkets, self.pointers_to_parent = kaolin.ops.spc.unbatched_make_trinkets(self.vox_point_all_levels, self.pyramids[0], self.point_hierarchy_dual, self.pyramid_dual)
    self.n_vox = len(self.vox_point_all_levels)
    self.n_corners = len(self.point_hierarchy_dual)

    for level in range(self.n_level):
      vox_pts = self.get_level_quantized_points(level)
      corner_pts = self.get_level_corner_quantized_points(level)
      logging.info(f'level:{level}, vox_pts:{vox_pts.shape}, corner_pts:{corner_pts.shape}')

  def get_level_corner_quantized_points(self,level):
    start = self.pyramid_dual[...,1,level]
    num = self.pyramid_dual[...,0,level]
    return self.point_hierarchy_dual[start:start+num]

  def get_level_quantized_points(self,level):
    start = self.pyramids[...,1,level]
    num = self.pyramids[...,0,level]
    return self.vox_point_all_levels[start:start+num]


  def get_center_ids(self,x,level):
    '''Get ids with 0 starting from current level's first point
    '''
    import kaolin
    pidx = kaolin.ops.spc.unbatched_query(self.octree, self.exsum, x.float(), level, with_parents=False)
    return pidx


  def get_vox_size_at_level(self, level):
    return 2.0/(2**level)


  def draw(self,level, method='point'):
    import kaolin
    logging.info(f"level:{level}")
    vox_size = self.get_vox_size_at_level(level)

    if method=='point':
      corner_coords = self.get_level_corner_quantized_points(level)
      pts = corner_coords*vox_size - 1
      mesh = trimesh.points.PointCloud(pts.data.cpu().numpy().reshape(-1,3))
      return mesh


  def ray_trace(self,rays_o,rays_d,level,debug=False):
    """Octree is in normalized [-1,1] world coordinate frame
    'rays_o': ray origin in normalized world coordinate system
    'rays_d': (N,3) unit length ray direction in normalized world coordinate system
    'octree': spc
    @voxel_size: in the scale of [-1,1] space
    Return:
        ray_depths_in_out: traveling times, NOT the Z value; invalid will be zeros
    """
    from mycuda import common
    import kaolin

    ray_index, rays_pid, depth_in_out = kaolin.render.spc.unbatched_raytrace(self.octree,self.vox_point_all_levels,self.pyramids[0],self.exsum,rays_o,rays_d,level=level,return_depth=True,with_exit=True)
    if ray_index.size()[0] == 0:
      pdb.set_trace()
      print("[WARNING] batch has 0 intersections!!")
      ray_depths_in_out = torch.zeros((rays_o.shape[0],1,2))
      rays_pid = -torch.ones_like(rays_o[:, :1])
      rays_near = torch.zeros_like(rays_o[:, :1])
      rays_far = torch.zeros_like(rays_o[:, :1])
      return rays_near, rays_far, rays_pid, ray_depths_in_out

    intersected_ray_ids,counts = torch.unique_consecutive(ray_index,return_counts=True)
    max_intersections = counts.max().item()
    start_poss = torch.cat([torch.tensor([0], device=counts.device),torch.cumsum(counts[:-1],dim=0)],dim=0)

    ray_depths_in_out = common.postprocessOctreeRayTracing(ray_index.long().contiguous(),depth_in_out.contiguous(),intersected_ray_ids.long().contiguous(),start_poss.long().contiguous(), max_intersections, rays_o.shape[0])

    rays_far = ray_depths_in_out[:,:,1].max(dim=-1)[0].reshape(-1,1)
    rays_near = ray_depths_in_out[:,0,0].reshape(-1,1)

    return rays_near, rays_far, rays_pid, ray_depths_in_out


def make_yaml_dumpable(D):
  if isinstance(D, np.ndarray):
    return D.tolist()
  for d in D:
    if isinstance(D[d], dict) or isinstance(D[d], OrderedDict) or isinstance(D[d], defaultdict):
      D[d] = dict(D[d])
      D[d] = make_yaml_dumpable(D[d])
      continue
    if isinstance(D[d], np.ndarray):
      D[d] = D[d].tolist()
      continue
    if np.issubdtype(type(D[d]), int):
      D[d] = int(D[d])
      continue
    if np.issubdtype(type(D[d]), float):
      D[d] = float(D[d])
      continue
    if np.issubdtype(type(D[d]), str):
      D[d] = str(D[d])
      continue
    if isinstance(D[d], list):
      for i in range(len(D[d])):
        D[d][i] = make_yaml_dumpable(D[d][i])
      continue
  return dict(D)