The list below gives a summary of how the design criteria mentioned in
Section 2 are fulfilled in the Columbia Photographic Images and Photoreal-
istic Computer Graphics Dataset:
1. PIM of a reliable source: The Personal set are from the authors'
and Greenspun's personal collection, therefore the images are certain
to be photograph.
2. PRCG from reliable sources: Only PRCG from trustable websites
are downloaded. These trustable websites requires the PRCG to be
submitted with the name and the contact information of the creator.
3. The content match for PIM and PRCG: The content of images in
the PIM and PRCG sets are limited to natural scene, which is defined
as scenes commonly encountered by human. PIM shown in Figure 5
and PRCG shown in Figure 4 demonstrate the content match between
the PRCG set and the Personal Columbia set.
4. PIM with diversity in terms of the image content, the camera
models and the photographer styles: Although the Personal set
has PIM of reliable source, it has limited diversity in terms of the
image content, the camera model and the photographer style. This
shortcoming is overcome by having the Google set, which is obtained
from the Internet and is supposedly diverse in terms of the image
content, the camera model and the photographer style.
5. PRCG with diversity in terms of the image content and the
rendering techniques: Ensuring a variety of PRCG content cate-
gories such as human, animal, building, scenery, indoor, outdoor, ob-
jects and so on. The PRCG are also generated using different high-end
computer graphics techniques (e.g., global illumination, soft shadow-
ing, environment map) and software (e.g., SoftImage, Maya and so
on).
6. PRCG of high photorealism and from high-quality rendering
techniques: From the available rendering information of the PRCG
set, the PRCG are mainly generated with high-quality rendering effects
such as global illumination, caustics effect, simulation of camera depth
of field and so on. The PRCG are also subjectively evaluated to be
highly photorealistic.
7. For evaluating the robustness of the classifier with respect to
the image processing operations on PIM: The RAW format of
7