Skip to main content

Table 2 Classification errors (≈ EERs) resulting from experiments on our auto-encoders and related studies with different data types

From: Estimation of gait normality index based on point clouds through deep auto-encoder

Model

Training data

Data type

Classification error (4 test subjects) †

   

Per-frame

Segment

Entire seq.

HMM [21]

Normal only

Skeleton

-

0.335

0.250

One-class SVM [5]

Normal only

Silhouette

0.399

0.227

0.139

Binary SVM [5]

Normal + abnormal

Silhouette

0.104

0.157

0.139

HMM [23]

Normal only

Depth map

-

0.396

0.281

Cross-correlation [23]

Normal only

Silhouette

-

0.381

0.250

HMM + cross-correlation [23]

Normal only

Silhouette + depth map

-

0.377

0.218

(Our) Sigmoid

Normal only

Point cloud

0.332

0.264

0.250

(Our) Sigmoid + dropout

Normal only

Point cloud

0.328

0.261

0.250

(Our) Tanh

Normal only

Point cloud

0.298

0.158

0.111

(Our) Tanh + dropout

Normal only

Point cloud

0.289

0.136

0.111

(Our) Leaky ReLU

Normal only

Point cloud

0.326

0.125

0.028

(Our) Leaky ReLU + dropout

Normal only

Point cloud

0.296

0.103

0.028

(Our) Multi-network

Normal only

Point cloud

0.288

0.125

0.083

  1. †Our system was originally implemented in Mathematica [37]. The models without dropout provided better results compared with the ones performed by TensorFlow [1] in this table. This may be because of the underlying algorithm implementation
  2. The italic values indicate the best results in different evaluations