singleshotpose基于2D图像实现3D输出

代码地址：https://github.com/Microsoft/singleshotpose
论文地址：Real-Time Seamless Single Shot 6D Object Pose Prediction

五、训练

python3 train.py --datacfg kitti/kitti.data --modelcfg kitti/yolo-pose.cfg --initweightfile cfg/darknet19_448.conv.23 --pretrain_num_epochs 15

训练之后的权重和损失会保存在backup里面

六、损失曲线

import numpy as np
import matplotlib.pyplot as plt
data = np.load('/root/work/21server/costs.npz', encoding='bytes', allow_pickle=True)
# print(data.files)
#['training_iters', 'testing_errors_angle', 'testing_iters', 'training_losses', 'testing_errors_pixel', 'testing_accuracies']
# print(data['testing_iters'])
# print(data['testing_iters'].shape)
# print(data['testing_accuracies'])
# print(data['testing_accuracies'].shape)
x1_data = data['testing_iters']
y1_data = data['testing_losses']
x2_data = data['training_iters']
y2_data = data['training_losses']

plt.figure(figsize=(6,6))
plt.plot(x1_data, y1_data,'r-',x2_data, y2_data,'b*')
plt.ylabel('loss')
plt.xlabel('iter')
plt.legend(['train', 'validation'], loc='upper right')
plt.savefig("loss.jpg")

六、测试结果可视化

1、新建backup/kitti/test/gt和backup/kiiti/test/pr，真实的角点坐标和预测的角点坐标会分别存储在这里；
2、valid.py的测试代码部分，稍微修改了一下；
3、测试命令：

python3 valid.py --datacfg kitti/kitti.data --modelcfg kitti/yolo-pose.cfg --weightfile backup/kitti/model_backup.weights

4、

import cv2
import os
import numpy as np


label_path = "/workspace/darknet/threedbbox/singleshotpose/LINEMOD/duck/labels/000000.txt"
pre_path = "/workspace/darknet/threedbbox/singleshotpose/backup/duck/test/pr/corners_0001.txt"
image_path = "/workspace/darknet/threedbbox/singleshotpose/LINEMOD/duck/JPEGImages/000000.jpg"

image = cv2.imread(image_path)
w = image.shape[1]
h = image.shape[0]
list_label1 = []
list_label2 = []
list_label3 = []
list_label4 = []

# with open(label_path, 'r') as f:
#     for line in f.readlines():
#         line_strs = line.split()
#         cls_id = line_strs[0]
#         for i in range(1,9):
#             x1 = int(float(line_strs[i*2+1])*w)
#             y1 = int(float(line_strs[i*2+2])*h)
#             list_label1.append((x1,y1))
#             list_label2.append((x1,y1))
#             cv2.circle(image,(x1,y1),3,(0,0,255),-1) # red
# print(list_label1)
# for label1 in list_label1:
#     for label2 in list_label2:
#         cv2.line(image, (label1), (label2), (0, 0, 255), 1)

# cv2.imwrite('000000_ground.jpg', image)

with open(pre_path, 'r') as f:
    for line in f.readlines():
        line_strs = line.split()
        x1 = int(float(line_strs[0]))
        y1 = int(float(line_strs[1]))
        list_label3.append((x1,y1))
        list_label4.append((x1,y1))
        cv2.circle(image,(x1,y1),3,(255,0,0),-1) # red
print(list_label3)
for label3 in list_label3:
    for label4 in list_label4:
        cv2.line(image, (label3), (label4), (0, 0, 255), 1)

cv2.imwrite('000000_pred.jpg', image)