assets/preprocessing/depth2normal.py

# common
import numpy as np
import open3d as o3d
import os
import glob
import argparse
import yaml
import cv2
from tqdm import tqdm


def config_setup():
	config = {}
	config["home_param"] = "<scene>/"
	config["depth_max"] = 10
	config["depth_min"] = 0.5
	return config

def load_depth(depth_path, config):
	# depth follows stanford-2D-3D-S (URL: http://buildingparser.stanford.edu/images/2D-3D-S_2017.pdf)
	# one unit of change in pixel intensity (e.g. a value from 45 to 46) corresponds to a 1/512 m change in depth (16 bit)
	depth_img = cv2.imread(depth_path, cv2.IMREAD_ANYDEPTH) / 512

	# depth is defined from depth_min to depth_max
	depth_img[depth_img > config["depth_max"]] = config["depth_max"]
	depth_img[depth_img < config["depth_min"]] = 0
	return depth_img

def equi2pcd(depth):
	# -----------------------------------------
	# image: [h,w,c], range:[0-255]
	# depth: [h,w], unit:meter
	# -----------------------------------------

	H,W = depth.shape

	# intrinsic parameter
	int_mtx = np.array([[max(H, W), 0, W/2], [0, max(H, W), H/2], [0, 0, 1]])
	if int_mtx.max() > 1:
		int_mtx[0, :] = int_mtx[0, :] / float(W)
		int_mtx[1, :] = int_mtx[1, :] / float(H)
	int_mtx_pix = int_mtx * np.array([[W], [H], [1.]])
	int_mtx_pix = int_mtx * np.array([[W], [H], [1.]])
	cam_param_pix_inv = np.linalg.inv(int_mtx_pix)
	k_00, k_02, k_11, k_12 = cam_param_pix_inv[0, 0], cam_param_pix_inv[0, 2], \
								cam_param_pix_inv[1, 1], cam_param_pix_inv[1, 2]

	# Convert from meshgrid of depth images to xyz 3D coordinates
	xyz = np.zeros((H*W,3))
	sx = np.arange(H).repeat(W)
	sy = np.arange(W)[None,:].repeat(H,axis=0).reshape(-1)
	sd = depth.reshape(-1)
	yaw = 2 * np.pi * ((sy+0.5) * k_00 + k_02)   # yaw:-π~πの範囲に変換
	pitch = 2 * np.pi * ((sx+0.5) * k_11 + k_12) # pitch:-π/2~π/2の範囲に変換
	xyz[:,0] = np.cos(pitch) * np.sin(yaw) * abs(sd)
	xyz[:,1] = np.sin(pitch) * abs(sd)
	xyz[:,2] = np.cos(pitch) * np.cos(yaw) * abs(sd)

	# Stored as Open3d pointcloud
	pcd = o3d.geometry.PointCloud()
	pcd.points = o3d.utility.Vector3dVector(xyz)

	return pcd

def pcd2normalimg(pcd, depth):
	# --------------------------------
	# Normals are calculated on pointcloud
	# --------------------------------
	H, W = depth.shape
	pcd.estimate_normals()
	normal = np.asarray(pcd.normals)
	normal = normal_align(normal, pcd)

	# stanford-2D-3D-S adapts left-handed system but open3d follows right-handed system
	normal[:,2] *= -1

	return normal
	
def normal_align(normal, pcd):
	# --------------------------------
	# All of the vector are aligned with the direction to the camera
	# --------------------------------
	points = np.asarray(pcd.points)
	vec2cam = np.array([0,0,0])[None,:].repeat(points.shape[0], axis=0) - points
	direction = np.sum(np.multiply(vec2cam, normal),axis=1) < 0
	normal[direction, :] *= -1
	return normal

def main():
	config = config_setup()
	print("home_path:", config["home_param"])

	save_folder_path = config["home_param"] + "normal/"
	if not os.path.exists(save_folder_path):
		os.mkdir(save_folder_path)

	# Search depth images defined as png
	depth_paths = sorted(glob.glob(config["home_param"] + "depth/*.png"))

	for idx, depth_path in tqdm(enumerate(depth_paths)):
		print("\n")
		print("depth file:", depth_path.split("/")[-1])
		depth = load_depth(depth_path, config)

		# To reduce the calculation costs
		H, W = (int(depth.shape[0]/4), int(depth.shape[1]/4)) 
		depth_img = cv2.resize(depth, (W, H), interpolation=cv2.INTER_NEAREST)
		
		pcd = equi2pcd(depth_img, config)
		normal = pcd2normalimg(pcd, depth_img)

		# To visualize the normal as colored pointcloud (It contains normal information itself)
		pcd.normals = o3d.utility.Vector3dVector(normal)
		pcd.colors = o3d.utility.Vector3dVector((normal+1)/2)
		# o3d.io.write_point_cloud(save_folder_path + f"{idx:03d}_"  + "equi_normal.ply", pcd)
		

		save_path = save_folder_path + f"{idx:03d}_"  + "equi_normal.png"
		print("output image:", save_path.split("/")[-1])

		# 8 bit value centered at 127.5 (-1~1 >> 0~255)
		normal_img = 127.5*(normal.reshape(H,W,3)+1.)
		
		# outlier, blank
		normal_img[depth_img<config["depth_min"], :] = [128,128,128]
		
		img_color = cv2.resize(cv2.cvtColor(normal_img.astype(np.uint8), cv2.COLOR_RGB2BGR), (depth.shape[1], depth.shape[0]), interpolation=cv2.INTER_NEAREST)
		cv2.imwrite(save_path, img_color)

if __name__ == "__main__":
	main()