DD-SVM
An image dataset of 2000 images
|
Image ID |
Category ID |
Category Name |
|
0 - 99 |
0 |
Africa People and Villages |
|
100 - 199 |
1 |
Beach |
|
200 - 299 |
2 |
Historical Buildings |
|
300 - 399 |
3 |
Buses |
|
400 - 499 |
4 |
Dinosaurs |
|
500 - 599 |
5 |
Elephants |
|
600 - 699 |
6 |
Flowers |
|
700 - 799 |
7 |
Horses |
|
800 - 899 |
8 |
Mountains and Glaciers |
|
900 - 999 |
9 |
Food |
|
1000 - 1099 |
10 |
Dogs |
|
1100 - 1199 |
11 |
Lizard |
|
1200 - 1299 |
12 |
Fashion |
|
1300 - 1399 |
13 |
Sunsets |
|
1400 - 1499 |
14 |
Cars |
|
1500 - 1599 |
15 |
Waterfall |
|
1600 - 1699 |
16 |
Antiques |
|
1700 - 1799 |
17 |
Battle Ships |
| 1800 - 1899 |
18 |
Skiing |
|
1900 - 1999 |
19 |
Dessert |
Image Features: MAT-file containing a 1x2000 cell array D and a 2000x1 array L (labels).
<D{i}, L(i)> is an image-label pair. Region features correspond to columns of D{i}.
The key step of LearnIPs in Algorithm 3.1 is line 5. It is implemented as emdd.m.
Please note that the negative label should be assigned as 0 instead of -1.
The rest of the algorithm should be straightforward to implement.
In Matlab, "help emdd" should give you more details.
The quasi-Newton search solver, argminh, is compiled for LINUX.