CSCLAB Image Database
In the
interest of advancing the field of computer vision, we provide
our object database as a freely available download to aid in building
and analyzing systems for image segmentation, stereo disparity, and
object recognition.
- The goal of this project was to assemble
scenes of
everyday objects in cluttered backgrounds with significant
occlusion. Using the 50 objects shown above, we assembled subsets
of 3 to 7 of them and captured approximately 500 scenes with
significant occlusion. The cameras were shown each object from
the same general viewpoint at all times, but subtle differences in
depth, rotation, camera lighting, etc. require a system to be invariant
to these changes which are ever-present in the visual world.
- In addition to the scenes, we provide scenes of all
10 backgrounds without objects, as well as scenes on every background
of each object by itself. For each scene we captued two images
using a pair of digital video cameras capable of capturing uncompressed
images. Following capture, we labeled all of the objects in
every scene and provide XML files detailing the objects present in each
scene and the a rectangle bound of their location in the scene.
The
images are all stored as PNG files to obtain lossless compression.
Here is an example of an image pair with superimposed labels:
(Click for a larger view)
Here are images of all 10
backgrounds:
Download
This download
is available for public use, but please do not distribute. Refer
anyone seeking the dataset to this webpage.
- Original dataset with
object labels (in xml format)
Also includes ground-truth images for
rectification and distortion removal:
origds.tgz (~1GB)
- Dataset corrected for camera lens distortion with
object labels
Also includes ground-truth images for
rectification:
lensds.tgz (~1GB)
- Dataset with images rectified for stereo processing with
object labels
rectds.tgz (~1GB)
Also provided in the downloads are rudimentary segmentation masks made from background segmentation.
Their accuracy is marginal, but they are still provided as a convenience.
Statistics
of the image dataset
Click
on this link to see the
statistical description of the image database.
However, expect the objects to
scale as much as 2x in their projection
on the images
|
