# coding: utf-8
import argparse
import torch
from torch.utils.data import DataLoader
import torch.nn as nn
import imageio
from spatial_network import build_SpatialNet, SpatialNet
from temporal_network import build_TemporalNet, TemporalNet
from smooth_network import build_SmoothNet, SmoothNet
import os
import numpy as np
import skimage
import cv2
import utils.torch_tps_transform as torch_tps_transform
import utils.torch_tps_transform_point as torch_tps_transform_point
from PIL import Image
import glob
import time
from torchvision.transforms import GaussianBlur
import grid_res
grid_h = grid_res.GRID_H
grid_w = grid_res.GRID_W
import matplotlib.pyplot as plt
plt.rcParams['axes.unicode_minus']=False
last_path = os.path.abspath(os.path.join(os.path.dirname("__file__"), os.path.pardir))
MODEL_DIR = os.path.join(last_path, 'full_model_tra')
def linear_blender(ref, tgt, ref_m, tgt_m, mask=False):
blur = GaussianBlur(kernel_size=(21,21), sigma=20)
r1, c1 = torch.nonzero(ref_m[0, 0], as_tuple=True)
r2, c2 = torch.nonzero(tgt_m[0, 0], as_tuple=True)
center1 = (r1.float().mean(), c1.float().mean())
center2 = (r2.float().mean(), c2.float().mean())
vec = (center2[0] - center1[0], center2[1] - center1[1])
ovl = (ref_m * tgt_m).round()[:, 0].unsqueeze(1)
ref_m_ = ref_m[:, 0].unsqueeze(1) - ovl
r, c = torch.nonzero(ovl[0, 0], as_tuple=True)
ovl_mask = torch.zeros_like(ref_m_).cuda()
proj_val = (r - center1[0]) * vec[0] + (c - center1[1]) * vec[1]
ovl_mask[ovl.bool()] = (proj_val - proj_val.min()) / (proj_val.max() - proj_val.min() + 1e-3)
mask1 = (blur(ref_m_ + (1-ovl_mask)*ref_m[:,0].unsqueeze(1)) * ref_m + ref_m_).clamp(0,1)
if mask: return mask1
mask2 = (1-mask1) * tgt_m
stit = ref * mask1 + tgt * mask2
return stit
def recover_mesh(norm_mesh, height, width):
#from [bs, pn, 2] to [bs, grid_h+1, grid_w+1, 2]
batch_size = norm_mesh.size()[0]
mesh_w = (norm_mesh[...,0]+1) * float(width) / 2.
mesh_h = (norm_mesh[...,1]+1) * float(height) / 2.
mesh = torch.stack([mesh_w, mesh_h], 2) # [bs,(grid_h+1)*(grid_w+1),2]
return mesh.reshape([batch_size, grid_h+1, grid_w+1, 2])
def get_rigid_mesh(batch_size, height, width):
ww = torch.matmul(torch.ones([grid_h+1, 1]), torch.unsqueeze(torch.linspace(0., float(width), grid_w+1), 0))
hh = torch.matmul(torch.unsqueeze(torch.linspace(0.0, float(height), grid_h+1), 1), torch.ones([1, grid_w+1]))
if torch.cuda.is_available():
ww = ww.cuda()
hh = hh.cuda()
ori_pt = torch.cat((ww.unsqueeze(2), hh.unsqueeze(2)),2) # (grid_h+1)*(grid_w+1)*2
ori_pt = ori_pt.unsqueeze(0).expand(batch_size, -1, -1, -1)
return ori_pt
def get_norm_mesh(mesh, height, width):
batch_size = mesh.size()[0]
mesh_w = mesh[...,0]*2./float(width) - 1.
mesh_h = mesh[...,1]*2./float(height) - 1.
norm_mesh = torch.stack([mesh_w, mesh_h], 3) # bs*(grid_h+1)*(grid_w+1)*2
return norm_mesh.reshape([batch_size, -1, 2]) # bs*-1*2
def test(args):
os.environ['CUDA_DEVICES_ORDER'] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
# define the network
spatial_net = SpatialNet()
temporal_net = TemporalNet()
smooth_net = SmoothNet()
if torch.cuda.is_available():
spatial_net = spatial_net.cuda()
temporal_net = temporal_net.cuda()
smooth_net = smooth_net.cuda()
#load the existing models if it exists
ckpt_list = glob.glob(MODEL_DIR + "/*.pth")
ckpt_list.sort()
if len(ckpt_list) == 3:
# load spatial warp model
spatial_model_path = MODEL_DIR + "/spatial_warp.pth"
spatial_checkpoint = torch.load(spatial_model_path)
spatial_net.load_state_dict(spatial_checkpoint['model'])
print('load model from {}!'.format(spatial_model_path))
# load temporal warp model
temporal_model_path = MODEL_DIR + "/temporal_warp.pth"
temporal_checkpoint = torch.load(temporal_model_path)
temporal_net.load_state_dict(temporal_checkpoint['model'])
print('load model from {}!'.format(temporal_model_path))
# load smooth warp model
smooth_model_path = MODEL_DIR + "/smooth_warp.pth"
smooth_checkpoint = torch.load(smooth_model_path)
smooth_net.load_state_dict(smooth_checkpoint['model'])
print('load model from {}!'.format(smooth_model_path))
else:
print('No checkpoint found!')
exit(0)
spatial_net.eval()
temporal_net.eval()
smooth_net.eval()
print("##################start testing#######################")
warp12_mesh1 = 0.
warp12_mesh2 = 0.
warp23_mesh1 = 0.
warp23_mesh2 = 0.
img1_list = []
img2_list = []
img3_list = []
video_path1 = args.video1_path
video_path2 = args.video2_path
video_path3 = args.video3_path
for i in range(2):
if i == 0:
video_frame_path1 = video_path1
video_frame_path2 = video_path2
else:
video_frame_path1 = video_path2
video_frame_path2 = video_path3
#define an online buffer (len == 7)
buffer_len = 7
tmotion_tensor_list1 = []
smotion_tensor_list1 = []
tmotion_tensor_list2 = []
smotion_tensor_list2 = []
# img name list
img1_name_list = glob.glob(os.path.join(video_frame_path1, '*.jpg'))
img2_name_list = glob.glob(os.path.join(video_frame_path2, '*.jpg'))
img1_name_list = sorted(img1_name_list)
img2_name_list = sorted(img2_name_list)
img1_tensor_list = []
img2_tensor_list = []
img1_hr_tensor_list = []
img2_hr_tensor_list = []
img_h = 360
img_w = 480
# load imgs
for k in range(0, len(img2_name_list)):
img1 = cv2.imread(img1_name_list[k])
# get high-resolution input
img1_hr = img1.astype(dtype=np.float32)
img1_hr = np.transpose(img1_hr, [2, 0, 1])
img1_hr_tensor = torch.tensor(img1_hr).unsqueeze(0)
img1_hr_tensor_list.append(img1_hr_tensor)
# get 360x480 input
img1 = cv2.resize(img1, (img_w, img_h))
img1 = img1.astype(dtype=np.float32)
img1 = np.transpose(img1, [2, 0, 1])
img1 = (img1 / 127.5) - 1.0
img1_tensor = torch.tensor(img1).unsqueeze(0)
img1_tensor_list.append(img1_tensor)
img2 = cv2.imread(img2_name_list[k])
# get high-resolution input
img2_hr = img2.astype(dtype=np.float32)
img2_hr = np.transpose(img2_hr, [2, 0, 1])
img2_hr_tensor = torch.tensor(img2_hr).unsqueeze(0)
img2_hr_tensor_list.append(img2_hr_tensor)
# get 360x480 input
img2 = cv2.resize(img2, (img_w, img_h))
img2 = img2.astype(dtype=np.float32)
img2 = np.transpose(img2, [2, 0, 1])
img2 = (img2 / 127.5) - 1.0
img2_tensor = torch.tensor(img2).unsqueeze(0)
img2_tensor_list.append(img2_tensor)
if i == 0:
img1_list = img1_hr_tensor_list
img2_list = img2_hr_tensor_list
else:
img3_list = img2_hr_tensor_list
start_time1 = time.time()
NOF = len(img2_name_list)
# motion estimation
for k in range(0, len(img2_name_list)):
# step 1: spatial warp
with torch.no_grad():
spatial_batch_out = build_SpatialNet(spatial_net, img1_tensor_list[k].cuda(), img2_tensor_list[k].cuda())
smotion1 = spatial_batch_out['motion1']
smotion2 = spatial_batch_out['motion2']
smotion_tensor_list1.append(smotion1)
smotion_tensor_list2.append(smotion2)
# step 2: temporal warp
with torch.no_grad():
temporal_batch_out1 = build_TemporalNet(temporal_net, img1_tensor_list)
temporal_batch_out2 = build_TemporalNet(temporal_net, img2_tensor_list)
tmotion_tensor_list1 = temporal_batch_out1['motion_list']
tmotion_tensor_list2 = temporal_batch_out2['motion_list']
print("fps (spatial & temporal warp):")
print(NOF/(time.time() - start_time1))
##############################################
############# data preparation ############
# converting tmotion (t-th frame) into tsmotion ( (t-1)-th frame )
rigid_mesh = get_rigid_mesh(1, img_h, img_w)
norm_rigid_mesh = get_norm_mesh(rigid_mesh, img_h, img_w)
smesh_list1 = []
smesh_list2 = []
tsmotion_list1 = []
tsmotion_list2 = []
for k in range(len(tmotion_tensor_list1)):
smotion1 = smotion_tensor_list1[k]
smesh1 = rigid_mesh + smotion1
smotion2 = smotion_tensor_list2[k]
smesh2 = rigid_mesh + smotion2
if k == 0:
tsmotion1 = smotion1.clone() * 0
tsmotion2 = smotion2.clone() * 0
else:
smotion1_1 = smotion_tensor_list1[k-1]
smesh1_1 = rigid_mesh + smotion1_1
tmotion1 = tmotion_tensor_list1[k]
tmesh1 = rigid_mesh + tmotion1
norm_smesh1_1 = get_norm_mesh(smesh1_1, img_h, img_w)
norm_tmesh1 = get_norm_mesh(tmesh1, img_h, img_w)
tsmesh1 = torch_tps_transform_point.transformer(norm_tmesh1, norm_rigid_mesh, norm_smesh1_1)
tsmotion1 = recover_mesh(tsmesh1, img_h, img_w) - smesh1
smotion2_1 = smotion_tensor_list2[k-1]
smesh2_1 = rigid_mesh + smotion2_1
tmotion2 = tmotion_tensor_list2[k]
tmesh2 = rigid_mesh + tmotion2
norm_smesh2_1 = get_norm_mesh(smesh2_1, img_h, img_w)
norm_tmesh2 = get_norm_mesh(tmesh2, img_h, img_w)
tsmesh2 = torch_tps_transform_point.transformer(norm_tmesh2, norm_rigid_mesh, norm_smesh2_1)
tsmotion2 = recover_mesh(tsmesh2, img_h, img_w) - smesh2
# append
smesh_list1.append(smesh1)
smesh_list2.append(smesh2)
tsmotion_list1.append(tsmotion1)
tsmotion_list2.append(tsmotion2)
# step 3: smooth warp
ori_mesh1 = 0
smooth_mesh1 = 0
delta_motion1 = 0
ori_mesh2 = 0
smooth_mesh2 = 0
delta_motion2 = 0
for k in range(len(tmotion_tensor_list1)-6):
#
tsmotion_sublist1 = tsmotion_list1[k:k+7]
tsmotion_sublist1[0] = tsmotion_sublist1[0] * 0
tsmotion_sublist2 = tsmotion_list2[k:k+7]
tsmotion_sublist2[0] = tsmotion_sublist2[0] * 0
with torch.no_grad():
smooth_batch_out = build_SmoothNet(smooth_net, tsmotion_sublist1, tsmotion_sublist2, smesh_list1[k:k+7], smesh_list2[k:k+7])
_ori_mesh1 = smooth_batch_out["ori_mesh1"]
_smooth_mesh1 = smooth_batch_out["smooth_mesh1"]
_ori_mesh2 = smooth_batch_out["ori_mesh2"]
_smooth_mesh2 = smooth_batch_out["smooth_mesh2"]
if k == 0:
ori_mesh1 = _ori_mesh1
smooth_mesh1 = _smooth_mesh1
ori_mesh2 = _ori_mesh2
smooth_mesh2 = _smooth_mesh2
else:
# for ref
ori_mesh1 = torch.cat((ori_mesh1, _ori_mesh1[:,-1,...].unsqueeze(1)), 1)
smooth_mesh1 = torch.cat((smooth_mesh1, _smooth_mesh1[:,-1,...].unsqueeze(1)), 1)
# for tgt
ori_mesh2 = torch.cat((ori_mesh2, _ori_mesh2[:,-1,...].unsqueeze(1)), 1)
smooth_mesh2 = torch.cat((smooth_mesh2, _smooth_mesh2[:,-1,...].unsqueeze(1)), 1)
print("fps (smooth warp):")
print(NOF/(time.time() - start_time1))
# get meshes
if i == 0:
warp12_mesh1 = smooth_mesh1 # 1, N, grid_h+1, grid_w+1, 2
warp12_mesh2 = smooth_mesh2
else:
warp23_mesh1 = smooth_mesh1
warp23_mesh2 = smooth_mesh2
########################################################################################
# resize the mesh to the original resolution
batch_size, _, img_h, img_w = img1_list[0].shape
warp12_mesh1 = torch.stack([warp12_mesh1[...,0]*img_w/480, warp12_mesh1[...,1]*img_h/360], 4)
warp12_mesh2 = torch.stack([warp12_mesh2[...,0]*img_w/480, warp12_mesh2[...,1]*img_h/360], 4)
warp23_mesh1 = torch.stack([warp23_mesh1[...,0]*img_w/480, warp23_mesh1[...,1]*img_h/360], 4)
warp23_mesh2 = torch.stack([warp23_mesh2[...,0]*img_w/480, warp23_mesh2[...,1]*img_h/360], 4)
# mesh alignment (we suppose warp12_mesh2 and warp23_mesh1 correspond to the same view)
offset = (warp12_mesh2 - warp23_mesh1).reshape(warp12_mesh2.shape[0], warp12_mesh2.shape[1], -1, 2) # bs, N, -1, 2
offset = torch.mean(offset, 2) # bs, N, 2
offset = offset.unsqueeze(2).unsqueeze(2) # bs, N, 1, 1, 2
# transform mesh coordinate
warp23_mesh1 = warp23_mesh1 + offset
warp23_mesh2 = warp23_mesh2 + offset
# find the middle mesh plane
middle_mesh = (warp12_mesh2 + warp23_mesh1)/2.
# predefined canvas
width_max1 = torch.max(warp12_mesh1[...,0])
width_max2 = torch.max(warp12_mesh2[...,0])
width_max3 = torch.max(warp23_mesh1[...,0])
width_max4 = torch.max(warp23_mesh2[...,0])
width_max = torch.maximum(width_max1, width_max2)
width_max = torch.maximum(width_max, width_max3)
width_max = torch.maximum(width_max, width_max4)
width_min1 = torch.min(warp12_mesh1[...,0])
width_min2 = torch.min(warp12_mesh2[...,0])
width_min3 = torch.min(warp23_mesh1[...,0])
width_min4 = torch.min(warp23_mesh2[...,0])
width_min = torch.minimum(width_min1, width_min2)
width_min = torch.minimum(width_min, width_min3)
width_min = torch.minimum(width_min, width_min4)
height_max1 = torch.max(warp12_mesh1[...,1])
height_max2 = torch.max(warp12_mesh2[...,1])
height_max3 = torch.max(warp23_mesh1[...,1])
height_max4 = torch.max(warp23_mesh2[...,1])
height_max = torch.maximum(height_max1, height_max2)
height_max = torch.maximum(height_max, height_max3)
height_max = torch.maximum(height_max, height_max4)
height_min1 = torch.min(warp12_mesh1[...,1])
height_min2 = torch.min(warp12_mesh2[...,1])
height_min3 = torch.min(warp23_mesh1[...,1])
height_min4 = torch.min(warp23_mesh2[...,1])
height_min = torch.minimum(height_min1, height_min2)
height_min = torch.minimum(height_min, height_min3)
height_min = torch.minimum(height_min, height_min4)
out_width = width_max - width_min
out_height = height_max - height_min
print("predefined canvas")
print(out_width)
print(out_height)
warp12_mesh1 = torch.stack([warp12_mesh1[...,0]-width_min, warp12_mesh1[...,1]-height_min], 4)
warp12_mesh2 = torch.stack([warp12_mesh2[...,0]-width_min, warp12_mesh2[...,1]-height_min], 4)
warp23_mesh1 = torch.stack([warp23_mesh1[...,0]-width_min, warp23_mesh1[...,1]-height_min], 4)
warp23_mesh2 = torch.stack([warp23_mesh2[...,0]-width_min, warp23_mesh2[...,1]-height_min], 4)
middle_mesh = torch.stack([middle_mesh[...,0]-width_min, middle_mesh[...,1]-height_min], 4)
warp12_mesh1_framelist = []
warp23_mesh2_framelist = []
for i in range(middle_mesh.shape[1]):
norm_warp12_mesh1 = get_norm_mesh(warp12_mesh1[:,i,...], out_height, out_width)
norm_warp12_mesh2 = get_norm_mesh(warp12_mesh2[:,i,...], out_height, out_width)
norm_warp23_mesh1 = get_norm_mesh(warp23_mesh1[:,i,...], out_height, out_width)
norm_warp23_mesh2 = get_norm_mesh(warp23_mesh2[:,i,...], out_height, out_width)
norm_middle_mesh = get_norm_mesh(middle_mesh[:,i,...], out_height, out_width)
norm_warp12_mesh1 = torch_tps_transform_point.transformer(norm_warp12_mesh1, norm_warp12_mesh2, norm_middle_mesh)
warp12_mesh1_frame = recover_mesh(norm_warp12_mesh1, out_height, out_width)
norm_warp23_mesh2 = torch_tps_transform_point.transformer(norm_warp23_mesh2, norm_warp23_mesh1, norm_middle_mesh)
warp23_mesh2_frame = recover_mesh(norm_warp23_mesh2, out_height, out_width)
warp12_mesh1_framelist.append(warp12_mesh1_frame)
warp23_mesh2_framelist.append(warp23_mesh2_frame)
warp12_mesh1 = torch.stack(warp12_mesh1_framelist, 1)
warp23_mesh2 = torch.stack(warp23_mesh2_framelist, 1)
print(warp12_mesh1.shape)
# new canvas
width_max1 = torch.max(warp12_mesh1[...,0])
width_max2 = torch.max(middle_mesh[...,0])
width_max3 = torch.max(warp23_mesh2[...,0])
width_max = torch.maximum(width_max1, width_max2)
width_max = torch.maximum(width_max, width_max3)
width_min1 = torch.min(warp12_mesh1[...,0])
width_min2 = torch.min(middle_mesh[...,0])
width_min3 = torch.min(warp23_mesh2[...,0])
width_min = torch.minimum(width_min1, width_min2)
width_min = torch.minimum(width_min, width_min3)
height_max1 = torch.max(warp12_mesh1[...,1])
height_max2 = torch.max(middle_mesh[...,1])
height_max3 = torch.max(warp23_mesh2[...,1])
height_max = torch.maximum(height_max1, height_max2)
height_max = torch.maximum(height_max, height_max3)
height_min1 = torch.min(warp12_mesh1[...,1])
height_min2 = torch.min(middle_mesh[...,1])
height_min3 = torch.min(warp23_mesh2[...,1])
height_min = torch.minimum(height_min1, height_min2)
height_min = torch.minimum(height_min, height_min3)
out_width = width_max - width_min
out_height = height_max - height_min
print("new canvas")
print(out_width)
print(out_height)
batch_size, _, img_h, img_w = img1_list[0].shape
print(img2_list[0].shape)
rigid_mesh = get_rigid_mesh(batch_size, img_h, img_w)
norm_rigid_mesh = get_norm_mesh(rigid_mesh, img_h, img_w)
stable_list = []
print("warping and blending")
# warp
for i in range(warp12_mesh1.shape[1]):
mesh1 = warp12_mesh1[:,i,:,:,:]
mesh_trans1 = torch.stack([mesh1[...,0]-width_min, mesh1[...,1]-height_min], 3)
norm_mesh1 = get_norm_mesh(mesh_trans1, out_height, out_width)
img1 = img1_list[i].cuda()
mesh2 = middle_mesh[:,i,:,:,:]
mesh_trans2 = torch.stack([mesh2[...,0]-width_min, mesh2[...,1]-height_min], 3)
norm_mesh2 = get_norm_mesh(mesh_trans2, out_height, out_width)
img2 = img2_list[i].cuda()
mesh3 = warp23_mesh2[:,i,:,:,:]
mesh_trans3 = torch.stack([mesh3[...,0]-width_min, mesh3[...,1]-height_min], 3)
norm_mesh3 = get_norm_mesh(mesh_trans3, out_height, out_width)
img3 = img3_list[i].cuda()
if args.fusion_mode == 'AVERAGE':
img_warp = torch_tps_transform.transformer(torch.cat([img1, img2, img3], 0), torch.cat([norm_mesh1, norm_mesh2, norm_mesh3], 0), torch.cat([norm_rigid_mesh, norm_rigid_mesh, norm_rigid_mesh], 0), (out_height.int(), out_width.int()), mode = args.warp_mode)
img12_fusion = img_warp[0] * (img_warp[0]/ (img_warp[0]+img_warp[1]+1e-6)) + img_warp[1] * (img_warp[1]/ (img_warp[0]+img_warp[1]+1e-6))
fusion = img12_fusion * (img12_fusion/ (img12_fusion+img_warp[2]+1e-6)) + img_warp[2] * (img_warp[2]/ (img12_fusion+img_warp[2]+1e-6))
else:
mask = torch.ones_like(img1[:,0,...].unsqueeze(1)).cuda()
img1 = torch.cat([img1, mask], 1)
img2 = torch.cat([img2, mask], 1)
img3 = torch.cat([img3, mask], 1)
img_warp = torch_tps_transform.transformer(torch.cat([img1, img2, img3], 0), torch.cat([norm_mesh1, norm_mesh2, norm_mesh3], 0), torch.cat([norm_rigid_mesh, norm_rigid_mesh, norm_rigid_mesh], 0), (out_height.int(), out_width.int()), mode = args.warp_mode)
mask1 = img_warp[0,3,...].unsqueeze(0).unsqueeze(0)
mask2 = img_warp[1,3,...].unsqueeze(0).unsqueeze(0)
mask3 = img_warp[2,3,...].unsqueeze(0).unsqueeze(0)
img12_fusion = linear_blender(img_warp[0,0:3,...].unsqueeze(0), img_warp[1,0:3,...].unsqueeze(0), mask1, mask2)
mask12 = mask1 + mask2 - mask1*mask2
fusion = linear_blender(img12_fusion, img_warp[2,0:3,...].unsqueeze(0), mask12, mask3)
fusion = fusion[0]
stable_list.append(fusion.cpu())
print("begin to write into video")
# save video
save_path = '../out.mp4'
fourcc = cv2.VideoWriter_fourcc('m', 'p', '4', 'v')
fps = 30
media_writer = cv2.VideoWriter(save_path, fourcc, fps, (out_width.int(), out_height.int()))
for k in range(len(stable_list)):
ave_fusion = stable_list[k].cpu().numpy().transpose(1,2,0)
media_writer.write(ave_fusion.astype(np.uint8 ))
media_writer.release()
print("##################end testing#######################")
if __name__=="__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', type=str, default='0')
# the path to load input videos
# Note: video1 should overlap with video2, and video2 should overlap with video3
parser.add_argument('--video1_path', type=str, default='/opt/data/private/nl/Data/Tra-Dataset2/case5_2/video1/')
parser.add_argument('--video2_path', type=str, default='/opt/data/private/nl/Data/Tra-Dataset2/case5_2/video2/')
parser.add_argument('--video3_path', type=str, default='/opt/data/private/nl/Data/Tra-Dataset2/case5_3/video2/')
# optional parameter: 'NORMAL' or 'FAST'
# FAST: use F.grid_sample to interpolate. It's fast, but may produce thin black boundary.
# NORMAL: use our implemented interpolation function. It's a bit slower, but avoid the black boundary.
parser.add_argument('--warp_mode', type=str, default='NORMAL') # optional parameter: 'Normal' or 'Fast'
# optional parameter: 'AVERAGE' or 'LINEAR'
# AVERAGE: faster but more artifacts
# LINEAR: slower but less artifacts
parser.add_argument('--fusion_mode', type=str, default='LINEAR')
print('<==================== Loading data ===================>\n')
args = parser.parse_args()
print(args)
test(args) 这段代码能够用来进行图像的拼接吗
超级会员免费看
订阅专栏 解锁全文
扫一扫
56
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
