AMAT Experiment

Proposal Experiement

Statistics on First 100 Images Pascal 2007 (AMAT with resize 0.33)

	Mean Average Best overlap for the box-based locations	Average Number of Boxes
Felzenszwalb	0.759	922.49
AMAT (w_s = e-4)	0.677	320.56
AMAT (w_s = e-3)	0.685	329.69
AMAT (w_s = e-2)	0.609	169.74

Statistics on First 100 Images Pascal 2007 (AMAT without resize)

	Mean Average Best overlap for the box-based locations	Average Number of Boxes
Felzenszwalb	0.759	922.49
AMAT (w_s = e-4)
AMAT (w_s = e-3)	0.728	790.00
AMAT (w_s = e-2)	0.705	426.36

Image Classification Task

Caltech101 dataset: 3 categories (67 images per category)

AlexNet Features + SVM (Trained on the original images)

	Accuracy
original	1.00
AMAT (w_s = e-4)	0.33
AMAT (w_s = e-3)	0.33
AMAT (w_s = e-2)	0.33

SURF Features + SVM (Trained on the original images)

	Accuracy
original	0.89
AMAT (w_s = e-4)	0.82
AMAT (w_s = e-3)	0.69
AMAT (w_s = e-2)	0.55

AlexNet Features + SVM (Trained on the reconstructed images)

	Accuracy
original	1.00
AMAT (w_s = e-4)	0.5461
AMAT (w_s = e-3)	0.4610
AMAT (w_s = e-2)	0.5674

SURF Features + SVM (Trained on the reconstructed images)

	Accuracy
original	0.89
AMAT (w_s = e-4)	0.90
AMAT (w_s = e-3)	0.89
AMAT (w_s = e-2)	0.83