MinkowskiEngine语义分割

MinkowskiEngine语义分割

要运行示例，请安装Open3D与PIP安装open3d-python。

cd /path/to/MinkowskiEngine

python -m examples.indoor

细分酒店房间

运行示例时，将看到一个旅馆房间和房间的语义分割。运行示例时，以交互方式旋转可视化效果。

首先，加载数据并体素化（量化）数据。调用MinkowskiEngine.utils.sparse_quantize进行体素化。

pcd = o3d.read_point_cloud(file_name)

coords = np.array(pcd.points)

feats = np.array(pcd.colors)

 

quantized_coords = np.floor(coords / voxel_size)

inds = ME.utils.sparse_quantize(quantized_coords)

准备体素化的坐标和特征后，应用MinkowskiEngine.SparseTensor将其包裹起来。此前，通过调用MinkowskiEngine.utils.sparse_collate来创建批处理。此函数采用一组坐标和特征并将其连接起来。还将批处理索引附加到坐标。最后，通过从颜色中减去0.5，对特征进行伪归一化。

# Create a batch, this process is done in a data loader during training in parallel.

batch = [load_file(config.file_name, 0.02)]

coordinates_, featrues_, pcds = list(zip(*batch))

coordinates, features = ME.utils.sparse_collate(coordinates_, featrues_)

 

# Normalize features and create a sparse tensor

sinput = ME.SparseTensor(features - 0.5, coords=coordinates).to(device)

最后，将稀疏张量前馈到网络中并获得预测。

soutput = model(sinput)

_, pred = soutput.F.max(1)

经过一些后处理。可以为标签着色，并排可视化输入和预测。

 

 

 运行示例后，权重会自动下载，并且权重目前是Scannet 3D分段基准测试中排名最高的算法。

有关更多详细信息，请参阅完整的室内细分示例。

import os

	import argparse
	import numpy as np
	from urllib.request import urlretrieve
	try:
	import open3d as o3d
	except ImportError:
	raise ImportError('Please install open3d with `pip install open3d`.')
	import torch
	import MinkowskiEngine as ME
	from examples.minkunet import MinkUNet34C
	from examples.common import Timer
	# Check if the weights and file exist and download
	if not os.path.isfile('weights.pth'):
	print('Downloading weights and a room ply file...')
	urlretrieve("http://cvgl.stanford.edu/data2/minkowskiengine/weights.pth",
	'weights.pth')
	urlretrieve("http://cvgl.stanford.edu/data2/minkowskiengine/1.ply", '1.ply')
	parser = argparse.ArgumentParser()
	parser.add_argument('--file_name', type=str, default='1.ply')
	parser.add_argument('--weights', type=str, default='weights.pth')
	parser.add_argument('--use_cpu', action='store_true')
	CLASS_LABELS = ('wall', 'floor', 'cabinet', 'bed', 'chair', 'sofa', 'table',
	'door', 'window', 'bookshelf', 'picture', 'counter', 'desk',
	'curtain', 'refrigerator', 'shower curtain', 'toilet', 'sink',
	'bathtub', 'otherfurniture')
	VALID_CLASS_IDS = [
	1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 16, 24, 28, 33, 34, 36, 39
	]
	SCANNET_COLOR_MAP = {
	0: (0., 0., 0.),
	1: (174., 199., 232.),
	2: (152., 223., 138.),
	3: (31., 119., 180.),
	4: (255., 187., 120.),
	5: (188., 189., 34.),
	6: (140., 86., 75.),
	7: (255., 152., 150.),
	8: (214., 39., 40.),
	9: (197., 176., 213.),
	10: (148., 103., 189.),
	11: (196., 156., 148.),
	12: (23., 190., 207.),
	14: (247., 182., 210.),
	15: (66., 188., 102.),
	16: (219., 219., 141.),
	17: (140., 57., 197.),
	18: (202., 185., 52.),
	19: (51., 176., 203.),
	20: (200., 54., 131.),
	21: (92., 193., 61.),
	22: (78., 71., 183.),
	23: (172., 114., 82.),
	24: (255., 127., 14.),
	25: (91., 163., 138.),
	26: (153., 98., 156.),
	27: (140., 153., 101.),
	28: (158., 218., 229.),
	29: (100., 125., 154.),
	30: (178., 127., 135.),
	32: (146., 111., 194.),
	33: (44., 160., 44.),
	34: (112., 128., 144.),
	35: (96., 207., 209.),
	36: (227., 119., 194.),
	37: (213., 92., 176.),
	38: (94., 106., 211.),
	39: (82., 84., 163.),
	40: (100., 85., 144.),
	}
	def load_file(file_name):
	pcd = o3d.io.read_point_cloud(file_name)
	coords = np.array(pcd.points)
	colors = np.array(pcd.colors)
	return coords, colors, pcd
	if __name__ == '__main__':
	config = parser.parse_args()
	device = torch.device('cuda' if (
	torch.cuda.is_available() and not config.use_cpu) else 'cpu')
	print(f"Using {device}")
	# Define a model and load the weights
	model = MinkUNet34C(3, 20).to(device)
	model_dict = torch.load(config.weights)
	model.load_state_dict(model_dict)
	model.eval()
	coords, colors, pcd = load_file(config.file_name)
	# Measure time
	with torch.no_grad():
	voxel_size = 0.02
	# Feed-forward pass and get the prediction
	in_field = ME.TensorField(
	features=torch.from_numpy(colors).float(),
	coordinates=ME.utils.batched_coordinates([coords / voxel_size], dtype=torch.float32),
	quantization_mode=ME.SparseTensorQuantizationMode.UNWEIGHTED_AVERAGE,
	minkowski_algorithm=ME.MinkowskiAlgorithm.SPEED_OPTIMIZED,
	device=device,
	)
	# Convert to a sparse tensor
	sinput = in_field.sparse()
	# Output sparse tensor
	soutput = model(sinput)
	# get the prediction on the input tensor field
	out_field = soutput.slice(in_field)
	logits = out_field.F
	_, pred = logits.max(1)
	pred = pred.cpu().numpy()
	# Create a point cloud file
	pred_pcd = o3d.geometry.PointCloud()
	# Map color
	colors = np.array([SCANNET_COLOR_MAP[VALID_CLASS_IDS[l]] for l in pred])
	pred_pcd.points = o3d.utility.Vector3dVector(coords)
	pred_pcd.colors = o3d.utility.Vector3dVector(colors / 255)
	pred_pcd.estimate_normals()
	# Move the original point cloud
	pcd.points = o3d.utility.Vector3dVector(
	np.array(pcd.points) + np.array([0, 5, 0]))
	# Visualize the input point cloud and the prediction
	o3d.visualization.draw_geometries([pcd, pred_pcd])