住宅の間取り図の Raster to Vector変換を行うFloor-Plan-Detection は GPUでの動作を前提?としているようですが、 内部で使用する CubiCasa5k の model_best_val_loss_var.pkl モデルは 200mb程度でしたので、CPUで動作させてみました。
参考url
実行環境
windows 11 + miniconda ( python 3.9 )
git clone や model , blender download
CONDA> git clone https://github.com/rbg-research/Floor-Plan-Detection CONDA> cd Floor-Plan-Detection download https://drive.google.com/uc?id=1gRB7ez1e4H7a9Y09lLqRuna0luZO5VRK as model_best_val_loss_var.pkl ( 203MB ) # vector化したものを 3D表示したい場合に使用します download https://www.blender.org/download/release/Blender2.93/blender-2.93.1-windows-x64.zip unzip blender-2.93.1-windows-x64.zip ( 203MB )
source の修正
github.com の Floor-Plan-Detection が 2021年頃の為でしょうか、 python 3.9 でエラーとなる箇所がありましたので、修正
utils/plotting.py
「ValueError: A colormap named "rooms_furu" is already registered.」の エラーとなりましたので、utils/plotting.py を以下のように修正しました
old) cmap3 = colors.ListedColormap(cpool, 'rooms_furu') new) cmap3 = colors.ListedColormap(cpool, 'rooms_icons')
config.py
vector化したものを 更に3D表示まではしないのであれば、変更は不要です
import os image_path = 'c:/Users/end0t/dev/test_floor_plan.png' target_path = './floorplan' # will export in two formats (.blend and .stl) program_path = os.getcwd() blender_install_path = program_path+"/blender-2.93.1-windows-x64/blender.exe" blender_script_path = program_path+"/floorplan_to_3dObject_in_blender.py" SR_scale = 2 SR_method = 'lapsrn' CubiCasa = True
utils/post_prosessing.py
scipy.stats.mode() の ver.1.9 での仕様変更の為
old) wall_width = stats.mode(widths).mode[0] new) wall_width = stats.mode(widths, keepdims=True).mode[0]
Floor-Plan-Detection cpu実行版
2D_FloorPlan_to_3D_CubiCasa_ftb.ipynb を改造したものです
import os import sys import numpy as np import matplotlib.pyplot as plt import matplotlib.image as mpimg import torch import torch.nn.functional as F import cv2 from torch.utils.data import DataLoader import FloorplanToSTL as stl import config from utils.FloorplanToBlenderLib import * from model import get_model from utils.loaders import FloorplanSVG, DictToTensor, Compose, RotateNTurns from utils.plotting import segmentation_plot, polygons_to_image, draw_junction_from_dict import utils.plotting from utils.post_prosessing import split_prediction, get_polygons, split_validation from mpl_toolkits.axes_grid1 import AxesGrid img_path = "c:/Users/end0t/dev/test_floor_plan.png" wall_height = 1 scale = 100 pkl_path = "model_best_val_loss_var.pkl" room_classes=["Background","Outdoor","Wall","Kitchen","Living Room","Bed Room", "Bath","Entry","Railing","Storage","Garage","Undefined"] icon_classes=["No Icon","Window","Door","Closet","Electrical Applience", "Toilet","Sink","Sauna Bench","Fire Place","Bathtub","Chimney"] def main(): """ segmentation 可視化で label毎に異なる色を割当てる color map作成。""" utils.plotting.discrete_cmap() rot:RotateNTurns = RotateNTurns() """ https://github.com/CubiCasa/CubiCasa5k/tree/master/floortrans/models """ model:hg_furukawa_original = get_model('hg_furukawa_original', 51) split = [21, len( room_classes ), len( icon_classes )] n_classes = split[0] + split[1] + split[2] # = 44 """ 最終出力層を上書き (class数44に合わせる) """ model.conv4_ = torch.nn.Conv2d( 256, n_classes, bias=True, kernel_size=1 ) """ 出力sizeを入力画像に合わせるための upsampling 層 修正 """ model.upsample = torch.nn.ConvTranspose2d( n_classes, n_classes, kernel_size=4, stride=4 ) checkpoint = torch.load( pkl_path, map_location='cpu' ) #CPU model.load_state_dict( checkpoint['model_state'] ) model.eval() # 推論modeへ切替え model.to('cpu') # CPUで動かす img = cv2.imread(img_path) # Create tensor for pytorch img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) # correct color channels img = 2 * (img / 255.0) - 1 # Image transformation to range (-1,1) # Move from (h,w,3)--->(3,h,w) as model input dimension is defined like this img = np.moveaxis(img, -1, 0) #img = torch.tensor([img.astype(np.float32)]) # .cuda() 削除 → CPU上に保持 img = torch.from_numpy(np.expand_dims(img.astype(np.float32), axis=0)) n_rooms = len( room_classes ) n_icons = len( icon_classes ) with torch.no_grad(): # 推論(≠学習)の為、勾配計算 無効化し memory節約 #Check if shape of image is odd or even size_check = np.array([img.shape[2], img.shape[3]]) % 2 height = img.shape[2] - size_check[0] width = img.shape[3] - size_check[1] img_size = (height, width) rotations = [ ( 0, 0), # 回転なし → 戻しなし ( 1, -1), # 90度回転 → -90度戻し ( 2, 2), # 180度回転 → 180度戻し (-1, 1) ] # -90度回転 → +90度戻し pred_count = len(rotations) prediction = torch.zeros([pred_count, n_classes, height, width]) # CPU上で生成 for i, r in enumerate( rotations ): forward, back = r rot_image = rot(img, 'tensor', forward) # 画像を正回転 pred = model(rot_image) # 予測 pred = rot(pred, 'tensor', back) # 結果の見た目を戻す # heatmap点の意味を戻す(例:icon向き) pred = rot(pred, 'points', back) # model出力 pred を指定のheight, widthにresize(補間) pred = F.interpolate(pred, size=(height, width), mode='bilinear', align_corners=True) prediction[i] = pred[0] # それぞれの回転での予測結果を平均 prediction = torch.mean(prediction, 0, True) rooms_pred = F.softmax(prediction[0, 21:21+12], 0).cpu().data.numpy() rooms_pred = np.argmax(rooms_pred, axis=0) icons_pred = F.softmax(prediction[0, 21+12:], 0).cpu().data.numpy() icons_pred = np.argmax(icons_pred, axis=0) heatmaps, rooms, icons = split_prediction(prediction, img_size, split) polygons, types, room_polygons, room_types = get_polygons((heatmaps, rooms, icons), 0.2, [1, 2]) # 壁polygon → 3D変換の準備 wall_polygon_numbers = [i for i, j in enumerate(types) if j['type'] == 'wall'] boxes = [] for i, j in enumerate(polygons): if i in wall_polygon_numbers: temp = [np.array([k]) for k in j] boxes.append(np.array(temp)) verts, faces, wall_amount = transform.create_nx4_verts_and_faces(boxes, wall_height, scale) # Create top walls verts verts = [] for box in boxes: verts.extend([transform.scale_point_to_vector(box, scale, 0)]) # create faces faces = [] for room in verts: temp = tuple(range(len(room))) faces.append([temp]) # 部屋やiconの画像を生成 pol_room_seg, pol_icon_seg = polygons_to_image(polygons, types, room_polygons, room_types, height, width ) # 画面表示 plt.figure(figsize=(12, 12)) ax = plt.subplot(1, 1, 1) ax.axis('off') rseg = ax.imshow(pol_room_seg, cmap='rooms', vmin=0, vmax=n_rooms - 0.1) cbar = plt.colorbar(rseg, ticks=np.arange(n_rooms) + 0.5, fraction=0.046, pad=0.01) cbar.ax.set_yticklabels(room_classes, fontsize=20) plt.tight_layout() plt.show() plt.figure(figsize=(12, 12)) ax = plt.subplot(1, 1, 1) ax.axis('off') iseg = ax.imshow(pol_icon_seg, cmap='icons', vmin=0, vmax=n_icons - 0.1) cbar = plt.colorbar(iseg, ticks=np.arange(n_icons) + 0.5, fraction=0.046, pad=0.01) cbar.ax.set_yticklabels(icon_classes, fontsize=20) plt.tight_layout() plt.show() # blender dataへ変換 stl.createFloorPlan(image_path = img_path, target_path = "floorplan", SR_Check=True ) if __name__ == '__main__': main()
実行結果
元画像
どこかの不動産サイトにあったものです
部屋や壁の抽出結果 (rooms)
ドアや窓の抽出結果 (icons)
rooms や icons の座標
stl.createFloorPlan()により、Data/以下に text fileが複数出力されます
- floor_faces.txt
- floor_verts.txt
- rooms_faces.txt
- rooms_verts.txt
- top_wall_faces.txt
- top_wall_verts.txt
- transform.txt
- wall_faces.txt
- wall_verts.txt
例えば、wall_verts.txt の抜粋は以下です。 壁面を縦?に見て出力されていますので、「wall_height = 1」の値があります
[[2.74, 11.71, 1], [2.74, 11.81, 1] [2.74, 11.71, 0], [2.74, 11.81, 0] ],
