k | Accuracy (%) |
---|---|
1 | 22.5 |
10 | 23.7 |
20 | 22.0 |
50 | 20.9 |
k | Step Size | Bin Size | Accuracy (%) |
---|---|---|---|
1 | 2 | 8 | 48.2 |
10 | 2 | 8 | 49.1 |
20 | 2 | 8 | 47.9 |
1 | 4 | 8 | 48.2 |
10 | 4 | 8 | 49.5 |
20 | 4 | 8 | 48.7 |
1 | 2 | 16 | 48.7 |
10 | 2 | 16 | 49.3 |
20 | 2 | 16 | 49.1 |
1 | 4 | 16 | 47.5 |
10 | 4 | 16 | 49.7 |
20 | 4 | 16 | 48.5 |
lambda | Step Size | Bin Size | Accuracy (%) |
---|---|---|---|
0.000001 | 2 | 8 | 66.1 |
0.000005 | 2 | 8 | 68.0 |
0.00001 | 2 | 8 | 67.6 |
0.00005 | 2 | 8 | 63.3 |
0.000001 | 4 | 8 | 66.0 |
0.000005 | 4 | 8 | 66.1 |
0.00001 | 4 | 8 | 66.5 |
0.00005 | 4 | 8 | 62.9 |
0.000001 | 2 | 16 | 62.1 |
0.000005 | 2 | 16 | 67.6 |
0.00001 | 2 | 16 | 66.0 |
0.00005 | 2 | 16 | 65.9 |
0.000001 | 4 | 16 | 61.5 |
0.000005 | 4 | 16 | 63.7 |
0.00001 | 4 | 16 | 63.5 |
0.00005 | 4 | 16 | 62.4 |
lambda | Step Size | Bin Size | Accuracy (%) |
---|---|---|---|
0.000001 | 4 | 16 | 70.1 |
0.000005 | 4 | 16 | 70.7 |
0.00001 | 4 | 16 | 68.0 |
0.00005 | 4 | 16 | 66.3 |
lambda | Step Size | Bin Size | Accuracy (%) |
---|---|---|---|
0.000001 | 4 | 16 | 67.1 |
0.000005 | 4 | 16 | 65.4 |
0.00001 | 4 | 16 | 63.0 |
0.00005 | 4 | 16 | 60.9 |
lambda | Step Size | Bin Size | Accuracy (%) |
---|---|---|---|
0.000001 | 4 | 16 | 67.9 |
0.000005 | 4 | 16 | 67.7 |
0.00001 | 4 | 16 | 63.0 |
0.00005 | 4 | 16 | 60.2 |
The self-similarity descriptors may work much better however, time constraint prevented me from experimenting with it. With the default settings it takes approx 37 seconds to calculate self-similarity descriptors for an image. Increasing patch size, descriptor radius and decreasing number of angular divisions within each sector I could get it down to 10 seconds per image but even this was prohibitive to run through all 3000 images.
lambda | Accuracy (%) |
---|---|
0.000001 | 71.6 |
0.000005 | 73.2 |
0.00001 | 70.9 |
0.00005 | 70.1 |
I attempted cross-validation wherein I take 100 samples per class with replacement for each of train and test set. Following are average accuracy and standard deviation observed over different number of iterations (Considering lambda = 0.000005, step size = 4 and bin size = 16 for bag of sifts:
iters | Avg. accuracy (%) | Standard deviation |
---|---|---|
1 | 59.5 | 0.19 |
2 | 60.8 | 0.18 |
5 | 59.7 | 0.14 |
10 | 60.34 | 0.14 |
Dictionary size | Accuracy (%) |
---|---|
10 | 51.9 |
20 | 60.8 |
50 | 61.7 |
100 | 65.34 |
200 | 68.0 |
500 | 69.34 |
1000 | 66.0 |
Category name | Accuracy | Sample training images | Sample true positives | False positives with true label | False negatives with wrong predicted label | ||||
---|---|---|---|---|---|---|---|---|---|
Kitchen | 0.630 | LivingRoom |
LivingRoom |
Bedroom |
InsideCity |
||||
Store | 0.610 | InsideCity |
InsideCity |
Coast |
LivingRoom |
||||
Bedroom | 0.450 | Industrial |
Office |
OpenCountry |
Office |
||||
LivingRoom | 0.380 | Bedroom |
Kitchen |
Suburb |
InsideCity |
||||
Office | 0.820 | Kitchen |
Store |
Industrial |
Bedroom |
||||
Industrial | 0.380 | Suburb |
InsideCity |
Office |
Store |
||||
Suburb | 0.950 | LivingRoom |
Street |
Forest |
Industrial |
||||
InsideCity | 0.460 | Street |
Street |
Store |
Kitchen |
||||
TallBuilding | 0.780 | Industrial |
Bedroom |
Forest |
Store |
||||
Street | 0.790 | Industrial |
TallBuilding |
Mountain |
TallBuilding |
||||
Highway | 0.810 | Industrial |
Street |
Coast |
Forest |
||||
OpenCountry | 0.530 | Coast |
Coast |
Street |
Forest |
||||
Coast | 0.770 | OpenCountry |
OpenCountry |
Highway |
Industrial |
||||
Mountain | 0.780 | Coast |
Store |
Coast |
Coast |
||||
Forest | 0.920 | Mountain |
TallBuilding |
Mountain |
Mountain |
||||
Category name | Accuracy | Sample training images | Sample true positives | False positives with true label | False negatives with wrong predicted label |
Category name | Accuracy | Sample training images | Sample true positives | False positives with true label | False negatives with wrong predicted label | ||||
---|---|---|---|---|---|---|---|---|---|
Kitchen | 0.750 | Store |
Office |
LivingRoom |
Bedroom |
||||
Store | 0.700 | Mountain |
Kitchen |
Kitchen |
Industrial |
||||
Bedroom | 0.540 | Industrial |
LivingRoom |
Industrial |
Industrial |
||||
LivingRoom | 0.530 | Bedroom |
Bedroom |
Bedroom |
Store |
||||
Office | 0.890 | Kitchen |
LivingRoom |
Kitchen |
Bedroom |
||||
Industrial | 0.470 | TallBuilding |
TallBuilding |
LivingRoom |
LivingRoom |
||||
Suburb | 0.960 | Industrial |
TallBuilding |
InsideCity |
LivingRoom |
||||
InsideCity | 0.700 | Coast |
Highway |
Industrial |
Kitchen |
||||
TallBuilding | 0.720 | Store |
Street |
Industrial |
InsideCity |
||||
Street | 0.750 | Highway |
TallBuilding |
Highway |
Mountain |
||||
Highway | 0.810 | Street |
Street |
InsideCity |
Mountain |
||||
OpenCountry | 0.700 | InsideCity |
Highway |
Coast |
Suburb |
||||
Coast | 0.680 | OpenCountry |
OpenCountry |
Mountain |
Highway |
||||
Mountain | 0.870 | LivingRoom |
OpenCountry |
Coast |
Forest |
||||
Forest | 0.910 | Bedroom |
Mountain |
Mountain |
Mountain |
||||
Category name | Accuracy | Sample training images | Sample true positives | False positives with true label | False negatives with wrong predicted label |