# Compute the error mean and std of the two-cam method
# Compare the error among three HPR models
import os
import pickle
import numpy as np
vit_large_err = {'target1': [], 'target2': [], 'target4': []}
vit_base_err = {'target1': [], 'target2': [], 'target4': []}
open_pose_err = {'target1': [], 'target2': [], 'target4': []}
data_path = '../data'
def angle_difference(vector_1, vector_2):
unit_vector_1 = vector_1 / np.linalg.norm(vector_1)
unit_vector_2 = vector_2 / np.linalg.norm(vector_2)
dot_product = np.dot(unit_vector_1, unit_vector_2)
if dot_product > 1:
dot_product = 1
angle = np.degrees(np.arccos(dot_product))
return angle
for SUBJECT_NAME in os.listdir(data_path):
subject_folder_path = os.path.join(data_path, SUBJECT_NAME)
if os.path.isfile(subject_folder_path):
continue
# frontal points: target 1 & target 2
front_folder_path = os.path.join(subject_folder_path, 'front')
with open(front_folder_path + '/ViTPose_large/final_target_normal_opt.pickle', 'rb') as f:
vit_large_pred = pickle.load(f)
with open(front_folder_path + '/ViTPose_base/final_target_normal_opt.pickle', 'rb') as f:
vit_base_pred = pickle.load(f)
with open(front_folder_path + '/OpenPose/final_target_normal_opt.pickle', 'rb') as f:
open_pose_pred = pickle.load(f)
with open(front_folder_path + '/two_cam_gt_normal.pickle', 'rb') as f:
gt = pickle.load(f)
vit_large_pred_tar1, vit_large_pred_tar2 = vit_large_pred['target1_normal'], vit_large_pred['target2_normal']
vit_base_pred_tar1, vit_base_pred_tar2 = vit_base_pred['target1_normal'], vit_base_pred['target2_normal']
open_pose_pred_tar1, open_pose_pred_tar2 = open_pose_pred['target1_normal'], open_pose_pred['target2_normal']
gt_tar1, gt_tar2 = gt['target1_normal'], gt['target2_normal']
side_folder_path = os.path.join(subject_folder_path, 'side')
with open(side_folder_path + '/ViTPose_large/final_target_normal_opt.pickle', 'rb') as f:
vit_large_pred = pickle.load(f)
with open(side_folder_path + '/ViTPose_base/final_target_normal_opt.pickle', 'rb') as f:
vit_base_pred = pickle.load(f)
with open(side_folder_path + '/OpenPose/final_target_normal_opt.pickle', 'rb') as f:
open_pose_pred = pickle.load(f)
with open(side_folder_path + '/two_cam_gt_normal.pickle', 'rb') as f:
gt = pickle.load(f)
vit_large_pred_tar4 = vit_large_pred['target4_normal']
vit_base_pred_tar4 = vit_base_pred['target4_normal']
open_pose_pred_tar4 = open_pose_pred['target4_normal']
gt_tar4 = gt['target4_normal']
vit_large_pred_tar_list = [vit_large_pred_tar1, vit_large_pred_tar2, vit_large_pred_tar4]
vit_base_pred_tar_list = [vit_base_pred_tar1, vit_base_pred_tar2, vit_base_pred_tar4]
open_pose_pred_tar_list = [open_pose_pred_tar1, open_pose_pred_tar2, open_pose_pred_tar4]
gt_tar_list = [gt_tar1, gt_tar2, gt_tar4]
for target, vit_large_pred_tar, vit_base_pred_tar, open_pose_pred_tar, gt_tar in zip(vit_large_err.keys(),
vit_large_pred_tar_list,
vit_base_pred_tar_list,
open_pose_pred_tar_list,
gt_tar_list):
vit_large_err[target].append(angle_difference(vit_large_pred_tar, gt_tar))
vit_base_err[target].append(angle_difference(vit_base_pred_tar, gt_tar))
# skip outlier for openpose target 4:
if target == 'target4' and (SUBJECT_NAME == 'charles_xu' or SUBJECT_NAME == 'jingyu_wu'):
continue
open_pose_err[target].append(angle_difference(open_pose_pred_tar, gt_tar))
err_dict = {'vit_large_err': vit_large_err, 'vit_base_err': vit_base_err, 'open_pose_err': open_pose_err}
with open('normal_error_dict_opt.pickle', 'wb') as f:
pickle.dump(err_dict, f)
# Compute difference mean and std
for target in vit_large_err.keys():
print("========================================================================")
for model_name, model_err in zip(err_dict.keys(), err_dict.values()):
print(f"{target}, {model_name}, mean = {np.mean(model_err[target]): .1f}")
print(f"{target}, {model_name}, std = {np.std(model_err[target]): .1f}")
