Function get_error_in_q

script/dm/pose_model.py:162–191

Convert Rotation matrix to quaternion, then calculate the location errors. original from PoseNet Paper

(args, dl, model, sample_size, device, batch_size=1)

# # pytorch
def get_error_in_q(args, dl, model, sample_size, device, batch_size=1):
    ''' Convert Rotation matrix to quaternion, then calculate the location errors. original from PoseNet Paper '''
    model.eval()
    
    results = np.zeros((sample_size, 2))
    results, vis_info = compute_error_in_q(args, dl, model, device, results, batch_size)
    median_result = np.median(results,axis=0)
    mean_result = np.mean(results,axis=0)

    # standard log
    print ('Median error {}m and {} degrees.'.format(median_result[0], median_result[1]))
    print ('Mean error {}m and {} degrees.'.format(mean_result[0], mean_result[1]))
    
    # timing log
    #print ('Avg execution time (sec): {:.3f}'.format(np.mean(time_spent)))

    # standard log2
    # num_translation_less_5cm = np.asarray(np.where(results[:,0]<0.05))[0]
    # num_rotation_less_5 = np.asarray(np.where(results[:,1]<5))[0]
    # print ('translation error less than 5cm {}/{}.'.format(num_translation_less_5cm.shape[0], results.shape[0]))
    # print ('rotation error less than 5 degree {}/{}.'.format(num_rotation_less_5.shape[0], results.shape[0]))
    # print ('results:', results)
    
    # save for direct-pn paper log
    # if 0:
    #   filename='Direct-PN+U_' + args.datadir.split('/')[-1] + '_result.txt'
    #   np.savetxt(filename, predict_pose)

    # visualize results
    # vis_pose(vis_info)

class EfficientNetB3(nn.Module):
    ''&#x27; EfficientNet-B3 backbone,