Upload recon1.py
Browse files- pifuhd/apps/recon1.py +182 -0
pifuhd/apps/recon1.py
ADDED
|
@@ -0,0 +1,182 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import io
|
| 2 |
+
import sys
|
| 3 |
+
import os
|
| 4 |
+
|
| 5 |
+
sys.path.insert(0, os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
|
| 6 |
+
ROOT_PATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
|
| 7 |
+
|
| 8 |
+
import time
|
| 9 |
+
import json
|
| 10 |
+
import numpy as np
|
| 11 |
+
import cv2
|
| 12 |
+
import random
|
| 13 |
+
import torch
|
| 14 |
+
import torch.nn as nn
|
| 15 |
+
from tqdm import tqdm
|
| 16 |
+
from torch.utils.data import DataLoader
|
| 17 |
+
import matplotlib.pyplot as plt
|
| 18 |
+
import matplotlib.cm as cm
|
| 19 |
+
import matplotlib
|
| 20 |
+
from numpy.linalg import inv
|
| 21 |
+
|
| 22 |
+
from lib.options import BaseOptions
|
| 23 |
+
from lib.mesh_util import reconstruction # save_obj_mesh_with_color
|
| 24 |
+
from lib.data import EvalWPoseDataset, EvalDataset
|
| 25 |
+
from lib.model import HGPIFuNetwNML, HGPIFuMRNet
|
| 26 |
+
from lib.geometry import index
|
| 27 |
+
|
| 28 |
+
from PIL import Image
|
| 29 |
+
from io import BytesIO
|
| 30 |
+
|
| 31 |
+
parser = BaseOptions()
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
def save_obj_mesh_with_color(output, verts, faces, colors=None):
|
| 35 |
+
# Ensure we're writing in binary mode for BytesIO
|
| 36 |
+
def write_data(f, data):
|
| 37 |
+
if isinstance(output, BytesIO):
|
| 38 |
+
f.write(data.encode('utf-8'))
|
| 39 |
+
else:
|
| 40 |
+
f.write(data)
|
| 41 |
+
|
| 42 |
+
if isinstance(output, str):
|
| 43 |
+
f = open(output, 'w')
|
| 44 |
+
elif isinstance(output, BytesIO):
|
| 45 |
+
f = output
|
| 46 |
+
else:
|
| 47 |
+
raise ValueError("output must be a filepath or BytesIO")
|
| 48 |
+
|
| 49 |
+
for i in range(verts.shape[0]):
|
| 50 |
+
if colors is not None:
|
| 51 |
+
line = f'v {verts[i, 0]} {verts[i, 1]} {verts[i, 2]} {colors[i, 0]} {colors[i, 1]} {colors[i, 2]}\n'
|
| 52 |
+
else:
|
| 53 |
+
line = f'v {verts[i, 0]} {verts[i, 1]} {verts[i, 2]}\n'
|
| 54 |
+
write_data(f, line)
|
| 55 |
+
|
| 56 |
+
for i in range(faces.shape[0]):
|
| 57 |
+
line = f'f {faces[i, 0] + 1} {faces[i, 1] + 1} {faces[i, 2] + 1}\n'
|
| 58 |
+
write_data(f, line)
|
| 59 |
+
|
| 60 |
+
if isinstance(output, str):
|
| 61 |
+
f.close()
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
def gen_mesh(res, net, cuda, data, thresh=0.5, use_octree=True, components=False):
|
| 66 |
+
image_tensor_global = data['img_512'].to(device=cuda)
|
| 67 |
+
image_tensor = data['img'].to(device=cuda)
|
| 68 |
+
calib_tensor = data['calib'].to(device=cuda)
|
| 69 |
+
|
| 70 |
+
net.filter_global(image_tensor_global)
|
| 71 |
+
net.filter_local(image_tensor[:, None])
|
| 72 |
+
|
| 73 |
+
try:
|
| 74 |
+
if net.netG.netF is not None:
|
| 75 |
+
image_tensor_global = torch.cat([image_tensor_global, net.netG.nmlF], 0)
|
| 76 |
+
if net.netG.netB is not None:
|
| 77 |
+
image_tensor_global = torch.cat([image_tensor_global, net.netG.nmlB], 0)
|
| 78 |
+
except Exception as e:
|
| 79 |
+
print(f"Error in netG filter: {e}")
|
| 80 |
+
|
| 81 |
+
b_min = data['b_min']
|
| 82 |
+
b_max = data['b_max']
|
| 83 |
+
|
| 84 |
+
try:
|
| 85 |
+
verts, faces, _, _ = reconstruction(
|
| 86 |
+
net, cuda, calib_tensor, res, b_min, b_max, thresh, use_octree=use_octree, num_samples=50000
|
| 87 |
+
)
|
| 88 |
+
verts_tensor = torch.from_numpy(verts.T).unsqueeze(0).to(device=cuda).float()
|
| 89 |
+
|
| 90 |
+
color = np.zeros(verts.shape)
|
| 91 |
+
interval = 50000
|
| 92 |
+
for i in range(len(color) // interval + 1):
|
| 93 |
+
left = i * interval
|
| 94 |
+
right = -1 if i == len(color) // interval else (i + 1) * interval
|
| 95 |
+
net.calc_normal(verts_tensor[:, None, :, left:right], calib_tensor[:, None], calib_tensor)
|
| 96 |
+
nml = net.nmls.detach().cpu().numpy()[0] * 0.5 + 0.5
|
| 97 |
+
color[left:right] = nml.T
|
| 98 |
+
|
| 99 |
+
obj_buffer = BytesIO()
|
| 100 |
+
save_obj_mesh_with_color(obj_buffer, verts, faces, color)
|
| 101 |
+
obj_buffer.seek(0) # Reset buffer pointer to the beginning
|
| 102 |
+
|
| 103 |
+
return obj_buffer
|
| 104 |
+
|
| 105 |
+
except Exception as e:
|
| 106 |
+
print(f"Error generating mesh: {e}")
|
| 107 |
+
return None
|
| 108 |
+
|
| 109 |
+
|
| 110 |
+
def recon(opt, use_rect=False):
|
| 111 |
+
state_dict_path = None
|
| 112 |
+
if opt.load_netMR_checkpoint_path is not None:
|
| 113 |
+
state_dict_path = opt.load_netMR_checkpoint_path
|
| 114 |
+
elif opt.resume_epoch < 0:
|
| 115 |
+
state_dict_path = f'{opt.checkpoints_path}/{opt.name}_train_latest'
|
| 116 |
+
opt.resume_epoch = 0
|
| 117 |
+
else:
|
| 118 |
+
state_dict_path = f'{opt.checkpoints_path}/{opt.name}_train_epoch_{opt.resume_epoch}'
|
| 119 |
+
|
| 120 |
+
start_id = opt.start_id
|
| 121 |
+
end_id = opt.end_id
|
| 122 |
+
|
| 123 |
+
cuda = torch.device(f'cuda:{opt.gpu_id}' if torch.cuda.is_available() else 'cpu')
|
| 124 |
+
|
| 125 |
+
if state_dict_path and os.path.exists(state_dict_path):
|
| 126 |
+
print(f'Resuming from {state_dict_path}')
|
| 127 |
+
state_dict = torch.load(state_dict_path, map_location=cuda)
|
| 128 |
+
print('Warning: opt is overwritten.')
|
| 129 |
+
dataroot = opt.dataroot
|
| 130 |
+
resolution = opt.resolution
|
| 131 |
+
results_path = opt.results_path
|
| 132 |
+
loadSize = opt.loadSize
|
| 133 |
+
|
| 134 |
+
opt = state_dict['opt']
|
| 135 |
+
opt.dataroot = dataroot
|
| 136 |
+
opt.resolution = resolution
|
| 137 |
+
opt.results_path = results_path
|
| 138 |
+
opt.loadSize = loadSize
|
| 139 |
+
else:
|
| 140 |
+
raise Exception(f'Failed to load state dict: {state_dict_path}')
|
| 141 |
+
|
| 142 |
+
test_dataset = EvalWPoseDataset(opt) if not use_rect else EvalDataset(opt)
|
| 143 |
+
print(f'Test data size: {len(test_dataset)}')
|
| 144 |
+
projection_mode = test_dataset.projection_mode
|
| 145 |
+
|
| 146 |
+
opt_netG = state_dict['opt_netG']
|
| 147 |
+
netG = HGPIFuNetwNML(opt_netG, projection_mode).to(device=cuda)
|
| 148 |
+
netMR = HGPIFuMRNet(opt, netG, projection_mode).to(device=cuda)
|
| 149 |
+
|
| 150 |
+
netG.eval()
|
| 151 |
+
netMR.load_state_dict(state_dict['model_state_dict'])
|
| 152 |
+
|
| 153 |
+
os.makedirs(opt.checkpoints_path, exist_ok=True)
|
| 154 |
+
os.makedirs(opt.results_path, exist_ok=True)
|
| 155 |
+
|
| 156 |
+
if start_id < 0:
|
| 157 |
+
start_id = 0
|
| 158 |
+
if end_id < 0:
|
| 159 |
+
end_id = len(test_dataset)
|
| 160 |
+
|
| 161 |
+
obj_buffers = []
|
| 162 |
+
with torch.no_grad():
|
| 163 |
+
print('Generating mesh (test) ...')
|
| 164 |
+
for i in tqdm(range(start_id, end_id)):
|
| 165 |
+
if i >= len(test_dataset):
|
| 166 |
+
break
|
| 167 |
+
|
| 168 |
+
test_data = test_dataset[i]
|
| 169 |
+
obj_buffer = gen_mesh(opt.resolution, netMR, cuda, test_data, components=opt.use_compose)
|
| 170 |
+
if obj_buffer:
|
| 171 |
+
obj_buffers.append(obj_buffer)
|
| 172 |
+
|
| 173 |
+
return obj_buffers
|
| 174 |
+
|
| 175 |
+
def reconWrapper(args=None, use_rect=False):
|
| 176 |
+
opt = parser.parse(args)
|
| 177 |
+
return recon(opt, use_rect)
|
| 178 |
+
|
| 179 |
+
if __name__ == '__main__':
|
| 180 |
+
buffers = reconWrapper()
|
| 181 |
+
if buffers:
|
| 182 |
+
print(f"Generated {len(buffers)} OBJ files in memory.")
|