|
News:
- CFP (pdf,
ps)
-
Register to obtain dataset.
Organizing Committee:
Leonid Sigal (Brown U)
Michael J. Black (Brown U)
Horst Haussecker (Intel)
Program Committee:
Ankur Agarwal (Microsoft Research)
Stefan Carlsson (KTH)
Trevor Darrell (MIT)
James Davis (UC Santa Cruz)
Larry Davis (U of Maryland)
David Fleet (U of Toronto)
David Forsyth (UIUC)
Pascal Fua (EPFL)
Horst Haussecker (Intel)
Daniel Huttenlocher (Cornell U)
Ram Nevatia (USC)
Deva Ramanan (TTI-C)
James Rehg (Georgia Tech)
Stan Sclaroff (Boston U)
Cristian Sminchisescu (TTI-C)
Philip Torr (Oxford Brookes)
Bill Triggs (INRIA)
Ying Wu (Northwestern U)
Ming-Hsuan Yang (Honda)
Invited Speakers:
David Fleet (U of Toronto)
Jessica Hodgins (CMU) [tentative]
Contacts:
E-mail:
ehum@cs.brown.edu
|
|
|
|
|
|
|
Submission Instructions: Submissions
should be made in PDF format on-line.
The submission web site is now operational. Authors should visit:
http://continue.cs.brown.edu/servlets/ehum07/continue.ss
and request an
account. Once the account is requested you will immediately get a
confirmation e-mail with login instructions.
When submitting a paper under
Author List: please write “Blind”. Do not list the actual names of
the authors! (The system we are using does not allow
blind submissions. This will hide the names of the authors from the
reviewers.) While the submission website cannot accept
supporting material, we
set up a separate FTP for that purpose.
Please
submit a single ZIP file containing all supplementary material
labeled with the paper number assigned by the submission system to:
ftp://ftp.cs.brown.edu/incoming/ehum2/. Reviewers will be advised to
look at these materials during the review process. Please read
the dataset and testing instructions bellow carefully.
|
|
|
Data: In addition to the
HumanEva-I dataset that we already made available to the community,
we are making an additional and much smaller dataset available for
reporting of results specifically for EHuM2 workshop. The
data can be downloaded (here)
and is referred to as HumanEva-II dataset. Unlike HumanEva-I the new
data was collected using hardware synchronized system with 4 color
video cameras and 12 Vicon motion capture cameras. Similar to the
HumanEva-I the motion capture was recorded at 120 FPS and video at
60 FPS. As a result we believe that synchronization and the quality
of the motion capture data was significantly improved. HumanEva-II
contains 2 test sequences and is distributed in similar format as
HumanEva-I with the same support code. As requested by participants
instead of distributing data in AVI format, the new data is
distributed as sequenced PNGs. This should make the data easier and
faster to use and as a result the new support code runs on Linux.
The two subjects in the HumanEva-II dataset correspond to subjects
S2 and S4 from the original HumanEva-I set, and we encourage the use
of HumanEva-I data for training if required. HumanEva-II dataset
also contains motion capture data for the first 3 frames of each of
the two test sequences. This motion capture data can be used to
extract the initial pose for tracking initialization or for deriving
skeletal measurements for the subject (when and if required).
Participants are very strongly encouraged
to use the datasets listed bellow for reporting of results. We are
aware that some approaches may not be able to report results on all
of the data, in such cases the priority for the reporting results is
defined from top-to-bottom of the list bellow. Hence, if the
instructions are followed, all submissions should in the very least
report relative 2D error in camera views C1 and C2 for the Set1
(frames 1-350) of HumanEva-II, Subject S2, Combo 1 sequence.
| Dataset |
Subject |
Action |
Trial |
Frames |
Results should be reported on the
following sub-sets |
| HumanEva-II |
S2 |
Combo |
1 |
1-1202 |
- Set1: Frames 1-350, Walking
- Set2: Frames 1-700, Walking + Jogging
- Set3: Frames 1-1202, Walking + Jogging +
Balancing
Note: Training motion capture data for this
subject is available in HumanEva-I. |
| HumanEva-II |
S4 |
Combo |
4 |
2-1258 |
- Set1: Frames 2-350, Walking
- Set2: Frames 2-700, Walking + Jogging
- Set3: Frames 2-1258, Walking + Jogging +
Balancing
Note: This subject does not have any publicly
available training data. |
| HumanEva-I* |
S1 |
Walking |
1 |
6 - 590 |
|
* For backward compatibility (with EHuM workshop) in addition to
the two sequences above, we also encourage reporting of
results on the HumanEva-I, Subject S1, Walking 1, Validation
sequence.
|
|
|
Reporting Results: One of the key
goals of this workshop is to compare the various approaches to
human motion and pose estimation. To this end we
very strongly encourage
the authors to use the following error metric for the approaches
mentioned. The evaluation on the test sequences above should be done
using the
on-line evaluation system.
| |
2D Approaches |
3D Approaches |
| Able to recover
(image/world) position of the body |
- Absolute 2D error (in pix)
- Relative 2D error (in pix)
|
- Absolute 3D error (in mm)
- Relative 3D error (in mm)
- Absolute 2D error (in pix)
- Relative 2D error (in pix)
|
| Not-able to recover
absolute
(world/image) position of the body |
- Relative 2D error (in pix)
|
- Relative 3D error (in mm)
- Relative 2D error (in pix)
|
The relative error is computed with respect to the pelvis (torsoDistal)
joint. The 2D error should be computed in image plane of cameras (C1)
and (C2). For 2D approaches this
means that views C1 and C2 should be used for pose estimation and
tracking. |
|
|
|
|
Assumptions: All assumptions
must be clearly stated. This will make it easier to classify
approaches and compare approaches of similar type. In particular, it
should be stated if the approach requires initialization, training,
knowledge of subjects measurements, etc. |
|
|
|
|
|
|
|
|
|
|
