MARPLE:
Matching Accuracy
Rated Pointwise
by Lenient Expert
In this page: * purpose * principle
* content
Purpose
This database was created as part of the IITiS-MECH
project, within the European CRIT-2
research programme. It contains a forensic expert's evaluation of local
similarities between fragments of dynamic toolmarks. The database is placed in
the public domain for the benefit of any researchers who may wish to develop
comparison techniques emulating the expert's judgment.
Principle
A number of striated toolmarks were photographed, and a straight line was placed
manually across the striae in each image. A program then projected a rectangular
area around the line into a 1D domain, to obtain a crossection or profile of the
toolmark. Another algorithm analyzed the profile to identify salient local
maxima of its derivative. A 32*32 pixel square window around each such salient
point is considered a feature.
All the features found in all toolmark images used in the study were compared
to each other by a forensic expert. The expert was presented two features at a
time and asked to declare them similar, dissimilar, or undecidable. He also had
the option of skipping some of the feature pairs.
The expert was instructed to rate similarity based on local feature
appearance only, and not to use extraneous information such as his own memory of
the entire toolmarks.
Content
The database contains can be downloaded as
complete file and
contains :
- 20 files named *.pic, containing microscopic images in
this format
- 3 text files named *.prz, in this
format. These files define which fragments were submitted to the expert
for comparison.
- 3 result files named *.out, containing expert's decisions in
this format. Every toolmark fragment was matched
against every other one defined by the same .prz file, and against itself
for completeness, so the number of expert judgements in an .out file is N2,
where N is the total number of fragments defined by the .prz file.
Notes on file formats:
Format of .pic files
Each file contains 512*512 greylevel pixels in a raw binary format. No header, 8
bits per pixel.
* purpose * principle
* back to content
Format of .prz files
These are plain ASCII text files. The format of each line of text is as follows:
FILENAME |
n |
xb |
yb |
xe |
ye |
p1 |
.... |
pn |
where:
- FILENAME
- name of a picture file (.pic) containing a toolmark
- n
- number of striae (stripes) identified in the toolmark
- xb, yb, xe, ye
- beginning and ending coordinates of a line representing a crossection of
the toolmark.
- p1 ... pn
- position of the center of each stria, measured along the crossection line.
Zero is located at (xb, yb).
* purpose * principle
* back to content
Format of .out files
Each < name >.out file corresponds to a < name >.prz file. It
contains an N*N table of char
(1-byte) values, where N is the total
number of fragments defined by the relevant .prz file (N = sum of all n values
in the .prz file).
| p1,1 |
... |
p1,n1 |
p2,1 |
... |
p2,n2 |
....... |
pm,1 |
... |
pm,nm |
p1,1 |
N*N cells with
values 0 through 4.
Values have following meanings:
0 |
not checked |
1 |
matching |
2 |
not matching |
3 |
don't know |
4 |
skipped |
|
... |
p1,n1 |
p2,1 |
... |
p2,n2 |
.......
.......
....... |
pm,1 |
... |
pm,nm |
(Note: The headers are added here for clarity. The file only contains the N*N
numbers, stored as bytes.)
* purpose * principle