Solutions to exercises in the book
Decision Forests in Computer Vision and Medical Image Analysis
A. Criminisi and J. Shotton
Springer 2013
Chapter 4: Classification Forests
- Excercise 4.1
| sw clas exp1_n2.txt /d 2 /t 8 | sw clas exp1_n2.txt /d 2 /t 200 /split linear |
 |
|
Using many trees and linear splits reduces artifacts. | |
- Excercise 4.2
| sw clas exp1_n4.txt /d 3 /t 200 /padx 2 /pady 2 | sw clas exp1_n4.txt /d 3 /t 200 /padx 2 /pady 2 /split linear |
|
|
|
The quality of the uncertainty away from training data is affected by the type of split function (weak learner). | |
- Excercise 4.3
| sw clas exp5_n2.txt /d 6 /t 200 | sw clas exp5_n2.txt /d 6 /t 200 /split linear |
 |
|
Using linear splits produces a possibly better separating surfaces. | |
- Excercise 4.4
| sw clas exp5_n4.txt /d 8 /t 200 /split linear | sw clas exp5_n4.txt /d 6 /t 200 /split linear |
|
|
|
Reducing the tree depth may cause underfitting and lower confidence. | |
- Excercise 4.5
| sw clas exp5_n4.txt /d 8 /t 400 /f 500 /split linear | sw clas exp5_n4.txt /d 8 /t 400 /f 3 /split linear |
|
|
|
Increasing randomness may reduce overall prediction confidence. | |
Chapter 5: Regression Forests
- Excercise 5.1
| sw regression exp2.txt /d 2 /t 100 | sw regression exp2.txt /d 6 /t 100 |
|
|
|
Large tree depth may lead to overfitting. | |
- Excercise 5.2
| sw regression exp3.txt /d 2 /t 400 | sw regression exp4.txt /d 2 /t 400 |
|
|
|
Larger training noise yields larger prediction uncertainty (wider pink region). | |
- Excercise 5.3
| sw regression exp7.txt /d 4 /t 200 | sw regression exp8.txt /d 4 /t 200 |
|
|
|
| sw regression exp9.txt /d 4 /t 200 | sw regression exp10.txt /d 4 /t 200 |
|
|
|
Non-linear curve fitting in diverse examples. Note the relatively smooth interpolation and extrapolation behaviour. | |
- Excercise 5.4
| sw regression exp11.txt /d 4 /t 200 | sw regression exp11.txt /d 6 /t 200 |
|
|
|
Single function regression does not capture the inherently ambiguous central region. But at least it returns an associated high uncertainty. | |
Chapter 6: Density Forests
- Excercise 6.1
| sw density exp1.txt /d 2 /t 3 | sw density exp1.txt /d 4 /t 3 |
|
|
|
Too deep trees may cause overfitting. | |
- Excercise 6.2
| sw density exp3.txt /d 4 /t 300 | sw density exp3.txt /d 6 /t 300 |
|
|
|
Too deep trees may cause overfitting. | |
- Excercise 6.3
| sw density exp7.txt /d 3 /t 300 | sw density exp7.txt /d 5 /t 300 |
|
|
|
Too deep trees may cause overfitting. | |
- Excercise 6.4
| sw density exp4.txt /d 3 /t 300 | sw density exp4.txt /d 5 /t 300 |
|
|
|
Too deep trees may cause overfitting. Some of the visible streaky artifacts are due to the use of axis-aligned weak learners. | |
Chapter 8: Semi-supervised Classification Forests
- Excercise 8.1
| sw ssclas exp1.txt /d 5 /t 100 | sw ssclas exp1.txt /d 5 /t 1 |
|
|
|
Note the larger uncertainty in the central region (left image). A single tree is always over-confident. | |
- Excercise 8.2
| sw ssclas exp1.txt /d 5 /t 200 /split linear | sw ssclas exp4.txt /d 5 /t 200 /split linear |
|
|
|
Adding further supervised data in the central region helps increase the prediction confidence . | |
- Excercise 8.3
| sw ssclas exp3.txt /d 5 /t 200 /split linear | sw ssclas exp3.txt /d 6 /t 200 /split linear |
|
|
|
Confidence decreases with training noise and increases with tree depth. | |
- Excercise 8.4
| sw ssclas exp5.txt /d 5 /t 200 /split linear | sw ssclas exp5.txt /d 5 /t 1 /split linear |
|
|
|
Single trees are over-confident. Using many random forests produces smooth uncertainty in the transition regions. | |
- Excercise 8.5
| sw ssclas exp9.txt /d 10 /t 200 /a 2 /split linear | sw ssclas exp10.txt /d 10 /t 200 /a 2 /split linear |
|
|
|
Adding the amount of supervision in regions of low confidence increases the prediction accuracy and the overall confidence. | |
