DenthDepth:深度估计(三维场景构建)——单目视觉挑战激光雷达

网上看到这个

MAXIEYE周圣砚：欲靠单目视觉挑战激光雷达

搜了一下，感觉应该就是这个工程，https://github.com/ialhashim/DenseDepth

宣传的检测效果

我网上搜一张图片，运行一下效果大致如下：

检测的准不准，自己判断了，我进行了一点点修改，只是把数据源从单张图片改为了从摄像头获取，

新建camera_test.py

import os
import glob
import argparse
import matplotlib
import cv2
import numpy as np
import matplotlib.pyplot as plt# Keras / TensorFlow
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '5'
from keras.models import load_model
from layers import BilinearUpSampling2D
from utils import predict, load_images, display_images, color_rescale
from matplotlib import pyplot as plt# Argument Parser
parser = argparse.ArgumentParser(description='High Quality Monocular Depth Estimation via Transfer Learning')
parser.add_argument('--model', default='nyu.h5', type=str, help='Trained Keras model file.')
args = parser.parse_args()# Custom object needed for inference and training
custom_objects = {'BilinearUpSampling2D': BilinearUpSampling2D, 'depth_loss_function': None}print('Loading model...')# Load model into GPU / CPU
model = load_model(args.model, custom_objects=custom_objects, compile=False)print('\nModel loaded ({0}).'.format(args.model))cap = cv2.VideoCapture(0)while True:# frame = cv2.imread("examples/626_image.png")ret, frame = cap.read()cv2.imshow("test", frame)frame = np.clip(np.asarray(frame[...,::-1], dtype=float) / 255, 0, 1)# Compute resultsoutputs = predict(model, frame[np.newaxis, :])cv2.imshow("depth", color_rescale(outputs[0]))if cv2.waitKey(1) & 0xFF == ord('q'):break

修改utils.py

import numpy as np
from PIL import Image
import cv2
import matplotlib.pyplot as plt
plasma = plt.get_cmap('plasma')def DepthNorm(x, maxDepth):return maxDepth / xdef predict(model, images, minDepth=10, maxDepth=1000, batch_size=2):# Support multiple RGBs, one RGB image, even grayscale if len(images.shape) < 3: images = np.stack((images,images,images), axis=2)if len(images.shape) < 4: images = images.reshape((1, images.shape[0], images.shape[1], images.shape[2]))# Compute predictionspredictions = model.predict(images, batch_size=batch_size)# Put in expected rangereturn np.clip(DepthNorm(predictions, maxDepth=maxDepth), minDepth, maxDepth) / maxDepthdef scale_up(scale, images):from skimage.transform import resizescaled = []for i in range(len(images)):img = images[i]output_shape = (scale * img.shape[0], scale * img.shape[1])scaled.append( resize(img, output_shape, order=1, preserve_range=True, mode='reflect', anti_aliasing=True ) )return np.stack(scaled)def load_images(image_files):loaded_images = []for file in image_files:x = np.clip(np.asarray(Image.open( file ), dtype=float) / 255, 0, 1)loaded_images.append(x)return np.stack(loaded_images, axis=0)def to_multichannel(i):if i.shape[2] == 3:return ii = i[:,:,0]print(i.shape)return np.stack((i,i,i), axis=2)def color_rescale(img):rescaled = img[:,:,0]print(np.min(rescaled))rescaled = rescaled - np.min(rescaled)print(np.max(rescaled))rescaled = rescaled / np.max(rescaled)return plasma(rescaled)[:,:,:3]def display_images(outputs, inputs=None, gt=None, is_colormap=True, is_rescale=True):import skimagefrom skimage.transform import resizeshape = (outputs[0].shape[0], outputs[0].shape[1], 3)all_images = []for i in range(outputs.shape[0]):imgs = []if isinstance(inputs, (list, tuple, np.ndarray)):x = to_multichannel(inputs[i])x = resize(x, shape, preserve_range=True, mode='reflect', anti_aliasing=True )imgs.append(x)if isinstance(gt, (list, tuple, np.ndarray)):x = to_multichannel(gt[i])x = resize(x, shape, preserve_range=True, mode='reflect', anti_aliasing=True )imgs.append(x)if is_colormap:imgs.append(color_rescale(outputs[i]))else:imgs.append(to_multichannel(outputs[i]))img_set = np.hstack(imgs)all_images.append(img_set)all_images = np.stack(all_images)return skimage.util.montage(all_images, multichannel=True, fill=(0,0,0))def save_images(filename, outputs, inputs=None, gt=None, is_colormap=True, is_rescale=False):montage =  display_images(outputs, inputs, is_colormap, is_rescale)im = Image.fromarray(np.uint8(montage*255))im.save(filename)def load_test_data(test_data_zip_file='nyu_test.zip'):print('Loading test data...', end='')import numpy as npfrom data import extract_zipdata = extract_zip(test_data_zip_file)from io import BytesIOrgb = np.load(BytesIO(data['eigen_test_rgb.npy']))depth = np.load(BytesIO(data['eigen_test_depth.npy']))crop = np.load(BytesIO(data['eigen_test_crop.npy']))print('Test data loaded.\n')return {'rgb':rgb, 'depth':depth, 'crop':crop}def compute_errors(gt, pred):thresh = np.maximum((gt / pred), (pred / gt))a1 = (thresh < 1.25   ).mean()a2 = (thresh < 1.25 ** 2).mean()a3 = (thresh < 1.25 ** 3).mean()abs_rel = np.mean(np.abs(gt - pred) / gt)rmse = (gt - pred) ** 2rmse = np.sqrt(rmse.mean())log_10 = (np.abs(np.log10(gt)-np.log10(pred))).mean()return a1, a2, a3, abs_rel, rmse, log_10def evaluate(model, rgb, depth, crop, batch_size=6, verbose=False):N = len(rgb)bs = batch_sizepredictions = []testSetDepths = []for i in range(N//bs):    x = rgb[(i)*bs:(i+1)*bs,:,:,:]# Compute resultstrue_y = depth[(i)*bs:(i+1)*bs,:,:]pred_y = scale_up(2, predict(model, x/255, minDepth=10, maxDepth=1000, batch_size=bs)[:,:,:,0]) * 10.0# Test time augmentation: mirror image estimatepred_y_flip = scale_up(2, predict(model, x[...,::-1,:]/255, minDepth=10, maxDepth=1000, batch_size=bs)[:,:,:,0]) * 10.0# Crop based on Eigen et al. croptrue_y = true_y[:,crop[0]:crop[1]+1, crop[2]:crop[3]+1]pred_y = pred_y[:,crop[0]:crop[1]+1, crop[2]:crop[3]+1]pred_y_flip = pred_y_flip[:,crop[0]:crop[1]+1, crop[2]:crop[3]+1]# Compute errors per image in batchfor j in range(len(true_y)):predictions.append(   (0.5 * pred_y[j]) + (0.5 * np.fliplr(pred_y_flip[j]))   )testSetDepths.append(   true_y[j]   )predictions = np.stack(predictions, axis=0)testSetDepths = np.stack(testSetDepths, axis=0)e = compute_errors(predictions, testSetDepths)if verbose:print("{:>10}, {:>10}, {:>10}, {:>10}, {:>10}, {:>10}".format('a1', 'a2', 'a3', 'rel', 'rms', 'log_10'))print("{:10.4f}, {:10.4f}, {:10.4f}, {:10.4f}, {:10.4f}, {:10.4f}".format(e[0],e[1],e[2],e[3],e[4],e[5]))return e

DenthDepth:深度估计(三维场景构建)——单目视觉挑战激光雷达相关推荐

单目深度估计方法：现状与前瞻
今天为大家推荐的是<中国图象图形学报>2019年第12期论文<单目深度估计技术进展综述>,该文由中国图象图形学学会成像探测与感知专委会组织,北京理工大学刘越教授等学者撰写,对国 ...
鱼眼深度估计！环视近场感知系列之几何预测
作者 | 王杰编辑 | 汽车人原文链接:https://zhuanlan.zhihu.com/p/584827087 点击下方卡片,关注"自动驾驶之心"公众号 ADAS巨卷干 ...
基于分段平面性的单目深度估计 P3Depth: Monocular Depth Estimation with a Piecewise Planarity Prior
P3Depth: Monocular Depth Estimation with a Piecewise Planarity Prior 面向可解释深度网络的单目深度估计 0 Abstract 单 ...
深度学习之单目深度估计：无监督学习篇
点击上方"3D视觉工坊",选择"星标" 干货第一时间送达作者:桔子毛 https://zhuanlan.zhihu.com/p/29968267 本文仅做学术 ...
纯视觉3D目标检测新SOTA！STS：基于Camera的深度估计新方法
点击下方卡片,关注"自动驾驶之心"公众号 ADAS巨卷干货,即可获取点击进入→自动驾驶之心技术交流群后台回复[领域综述]获取自动驾驶全栈近80篇综述论文! 论文链接:https ...
MyDLNote-360Camera：全景图像深度估计，结合 equirectangular 和 cubemap 两种映射（2020CVPR BiFuse）
BiFuse: Monocular 360◦ Depth Estimation via Bi-Projection Fusion [paper] [github] [project] Fig. 1. ...
【研究报告】从单目深度估计到单目三维场景重建-沈春华老师-VALSE Webinar 22-13（总第279期）
从单目深度估计到单目三维场景重建-沈春华老师-VALSE Webinar 22-13(总第279期) 报告总结 & 相关论文论文代码相关术语前言研究问题单目深度估计单目三维场景重建 ...
基于单目视觉深度估计的论文研究
单目视觉 From Big to Small: Multi-Scale Local Planar Guidance for Monocular Depth Estimation 作者:Jin Han ...
三维感知，这些干货足够了！(自动驾驶/三维重建/SLAM/点云/标定/深度估计/3D检测）...
人工智能渗入到越来越多领域,以计算机视觉为主的相关应用更是呈现爆发性增长,国内外资本对AI视觉兴趣只增不减,自动驾驶.工业视觉.AR/VR.测量测绘.移动机器人等领域涌现了大量独角兽公司,相关产品受到 ...

DenthDepth:深度估计(三维场景构建)——单目视觉挑战激光雷达

MAXIEYE周圣砚：欲靠单目视觉挑战激光雷达

DenthDepth:深度估计(三维场景构建)——单目视觉挑战激光雷达相关推荐

最新文章

热门文章