|
|||
|
File
Formats and Directory Structure Image Sets: |
Stool
Number Of Images: Total of 8201 images taken in 3
rooms: A401 (4423), A408 (2202), SH201 (1574) Image Size: 360 x 240 Comments: These images are frames
from a set of videos taken of a bar stool in a set of office rooms on a
handheld Sony camcorder. The
location of the stool in each frame is demarcated by a set of polygons. To label the location of the stool
over an entire video, we initialized a set of polygons to cover the stool as
it appeared in the first frame of the video—one polygon per leg,
cross-brace, and seat of the stool.
We then applied a 2-frame multiresolution Lucas-Kanade optical flow
estimator (cvCalcOpticalFlowPyrLK
in the OpenCV library) to estimate pixel flows between frames, and estimated
polygon motion based on the flows.
Polygon positions were manually adjusted at points of tracking
failure. We followed this
procedure for a total of 29 videos shot in 3 rooms: A401 (20 videos), A408 (3
videos), and SH201 (2 videos).
In each video, the cameraman keeps the stool in view and translates
and rotates around the stool. This
large image set allows the investigation of challenging problems: some of the
images are blurry, the stool changes scale significantly, and the change in
background between rooms is significant. Besides providing a large data set for wiry object
recognition, this image set provides benchmark data for a problem area that
(to my knowledge) has yet to be addressed: tracking complex-shaped wiry
non-human objects. Download: (copyright) All
rooms: [ stool.tar.gz , 127
MB] Individual rooms: [ a401_clips.tar.gz
, 21 MB ] [ a401_clips2.tar.gz
, 55 MB ] [ a408_clips.tar.gz
, 30 MB ] [ sh201_clips.tar.gz
, 20 MB ] |
||
WORD: Wiry
Object Recognition Database
