Compares ground truth normal maps and the ones predicted by FrameNet

framenet_normal_eval.py

'''
Compares ground truth normal maps and the ones predicted by FrameNet
'''

real_normal_maps_paths = [
'chair/test/chair_0890_001.jpg',
'chair/test/chair_0904_009.jpg',
'desk/test/desk_0201_004.jpg',
'desk/test/desk_0203_012.jpg',
'dresser/test/dresser_0201_005.jpg',
'dresser/test/dresser_0203_011.jpg',
'monitor/test/monitor_0466_002.jpg',
'monitor/test/monitor_0469_008.jpg'
]

# read both normal maps
import glob
import cv2
import numpy as np
pred_normal_maps = []
scale = 2.0/255.0
for file in sorted(glob.glob("framenet_pred/*pred.png")):
	# also convert bump mapping to to [-1, 1]
	pred_normal_maps.append(cv2.imread(file) * scale - 1)

real_normal_maps = []
masks = []
for path in real_normal_maps_paths:
	file = 'ModelNet40_SurfLinear/' + path
	orig_img = cv2.imread(file)
	# only consider pixels that correspond to an object
	resized = cv2.resize(orig_img, (320,240))
	mask = np.any(resized < [250, 250, 250], axis=-1)
	# only consider pixels that correspond to an object
	masks.append(mask)
	# linear mapping from [0,255] to [-1,1]
	real_normal_maps.append(resized * scale - 1)

# compare directions
for i in range(len(pred_normal_maps)):
	result = []
	for r in range(len(pred_normal_maps[i])):
		for c in range(len(pred_normal_maps[i][0])):
			dot_product = np.dot(pred_normal_maps[i][r][c], real_normal_maps[i][r][c])
			dot_product = np.clip(dot_product, -1, 1)
			angle = np.arccos(dot_product) * masks[i][r][c] / np.pi * 180
			result.append(angle)
	print(np.mean(result))