Eigen-Gradients for Traffic Sign Recognition

Table 1

General comparison with the methods reported in [2]. In the central column a brief description of the method is given: classification tool, image size, number of attributes per image, and features used for recognition.

Team	Method	Accuracy %
IDSIA [3]	CNN + MLP, image size 48×48, 8 attributes per pixel (18,432 attributes per image), color features	99.46
INI-RTCV	Human performance (best individual)	99.22
INI-RTCV	Human performance (average)	98.84
Sermanet [4]	CNN, image size 32×32, 3,072 attributes per image, color features	98.31
CAOR [5]	k-NN + Random Forests, image size 40×40, 1,568 attributes per image, HOG 2 features	96.14
INI-RTCV	LDA, image size 40×40, 1,568 attributes per image, HOG 2 features	95.68
INI-RTCV	LDA, image size 40×40, 1,568 attributes per image, HOG 1 features	93.18
INI-RTCV	LDA, image size 40×40, 2,916 attributes per image, HOG 3 features	92.34
Eigen-Gradients	MLP, image size 28×28, 140 attributes per image, gradient-orientation + PCA features	95.9