Buffy pose classes V 1.0
========================
Vittorio Ferrari, Manuel J. Marin-Jimenez and Andrew Zisserman
Description
~~~~~~~~~~~
Welcome to the Buffy pose classes dataset !
This package contains annotations for episodes 2 to 6 of the 5th season the TV show 'Buffy the vampire slayer'.
A pose class is characterized by a distinctive spatial configuration of the body parts.
For example, persons in the 'hips' pose have the hands on their hips, so that their arms stretch out in a typical
'diamond' shape. Pose classes are independent of gender, clothing, and other characteristic of a person.
Only the spatial configuration of the body parts matters.
This dataset is useful for testing pose retrieval systems, such as [1].
For a complete description of the recommended experimental setup and evaluation measure, please read [1].
Contents
~~~~~~~~
This package is organized in the following directories:
poses_gt_intervals:
15 ground-truth text files. One per episode and pose class: rest, hips and folded (as in the experiments of [1]).
The filenames indicate the episode and the pose. For example, the file 'buffy_s5e2_hips_gt.txt' contains
annotations for episode 2 and pose 'hips'. The ground-truth files indicate all frames containing an instance of
the pose class, as well as frames that contain an ambiguous pose (more details below). All frames not listed in
the ground-truth files do NOT contain an instance of the pose class. Therefore, the annotation is complete.
poses_gt_example_images:
5 directories containing sample frames for the different poses in the different episodes. The bounding-box returned
by the upper-body detector [4] is overlaid for the person carrying out the ground-truth pose.
shots:
shot boundaries for the 5 episodes (eg. 'buffy_s5e2_shots.txt'). Each line in such a file reports the first and the
last frame number of a shot. The subdirectories include example frames at shot boundaries, for synchronization purposes.
More precisely, the first and last frame of various shots are included. For example, image 'shots\buffy_s5e2\004525.jpg'
is the last frame of shot 99 (line #99 in file 'shots\buffy_s5e2_shots.txt') for episode 'buffy_s5e2'. Moreover, image
'shots\buffy_s5e2\004526.jpg' is the first frame of shot 100 (line #100 in file 'shots\buffy_s5e2_shots.txt').
This information is useful for checking whether the frames you extract from a DVD match our frame numbers (and to compute
the offset if they don't).
poses_gt_shots:
complementarily to the content of 'poses_gt_intervals', these files contain the list of shot numbers where each pose
class is found, along with a 'not_omit' flag (see description below). This is provided for convenience: the same information
could be derived by combining poses_gt_intervals with the shot boundary files in the directory 'shots'.
queries:
the subdirectories contain the query frames used in [1] organized by classes and episodes.
The additional directory 'BBs' include the bounding-boxes of the person carrying out the
query pose, as used in [1]. Each line in the files (eg. 'buffy_s5e2_hips_bbs.txt') has the following format:
<frame_number> <bb_x_min> <bb_x_max> <bb_y_min> <bb_y_max>
where each entry is a coordinate of the upper-body bounding-box (albeit the name, it covers
the head and shulders, not the whole upper-body).
Description of the ground-truth files
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Each file in the directory 'poses_gt_intervals' lists frame intervals in the following format.
Every line corresponds to a different instance of a pose class:
<initial_frame_i1> <final_frame_i1> <not_omit_i1>
<initial_frame_i2> <final_frame_i2> <not_omit_i2>
<initial_frame_i3> <final_frame_i3> <not_omit_i3>
. . .
. . .
Where:
- <initial_frame_iX>: the initial frame number of the X-th annotated instance.
- <final_frame_iX> the final frame number of the X-th annotated instance.
- <not_omit_iX> indicates whether to consider this instance in the retrieval evaluation or not.
If this flag is set to 1, then this is a clear instance of the pose, and should be included in
the evaluation. If the flag is set to 0, then this instance is ambiguous and should be omitted
from evaluation (see discussion in [1]).
Note that the evaluation in [1] only requires ground-truth annotation at the shot-level, i.e. it
only needs to know which shots contain a certain pose (i.e. as given in poses_gt_shots).
The annotation at the frame level in poses_gt_intervals gives an additional level of detail,
not needed for evaluation purposes.
Evaluation protocol
-------------------
The data in this package can be used to reproduce the 'query-mode' experiments in [1].
We briefly describe here how to evaluate the performance of a system for the pose class 'hips'
(the same can be done for any pose class):
1. given the first query frame for hips, the system must rank all shots in the 5 test videos
according to whether they contain the pose in the query frame. Any person in the shot can be doing the query pose,
in any clothing, background, illumination, and so on.
2. from the ranked list of shots a precision/recall plot should be derived in the standard manner,
and the average-precision computed. For this, use as ground-truth the information in 'poses_gt_shots'.
Notice how the query frame itself is contained in the 5 test videos. Therefore, you can expect the
system to place the shot containing the query frame at rank 1.
The procedure in points 1+2 must be repeated for all 7 query frames independently, i.e. using only one query frame at the time.
The final performance of the system is the mean average-precision over the 7 queries.
Shots marked as ambiguous in the ground-truth (i.e. not_omit = 0) should be ignored (i.e. removed entirely from the evaluation).
For a detailed description of the protocol we refer to [1].
References
~~~~~~~~~~
Experiments on this dataset first appeared in:
[1] Pose search: retrieving people using their pose
Ferrari, V. , Marin-Jimenez, M. and Zisserman, A.
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2009)
The pose estimator used in [1] was built on top of:
[2] Progressive Search Space Reduction for Human Pose Estimation
Ferrari, V., Marin-Jimenez, M. and Zisserman, A.
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2008)
[3] 2D Human Pose Estimation in TV Shows
Ferrari, V., Marin-Jimenez, M. and Zisserman, A.
Statistical and Geometrical Approaches to Visual Motion Analysis, proceedings of the Dagstuhl workshop (2009)
If you are interested in our frontal upper-body detector:
[4] http://www.robots.ox.ac.uk/~vgg/software/UpperBody/index.html
If you use this dataset in your publications, please cite [1].
Support
~~~~~~~
For any query/suggestion/complaint or simply to say you like/use this dataset, just drop us an email
ferrari@vision.ee.ethz.ch
mjmarin@uco.es
az@robots.ox.ac.uk
没有合适的资源?快使用搜索试试~ 我知道了~
Buffy pose 人类姿势图像数据.7z
共522个文件
jpg:454个
txt:47个
db:21个
1.该资源内容由用户上传,如若侵权请联系客服进行举报
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
版权申诉
0 下载量 93 浏览量
2022-07-13
13:09:36
上传
评论
收藏 12.56MB 7Z 举报
温馨提示
Buffy pose 人类姿势图像数据.7z
资源详情
资源评论
资源推荐
收起资源包目录
Buffy pose 人类姿势图像数据.7z (522个子文件)
Thumbs.db 318KB
Thumbs.db 271KB
Thumbs.db 242KB
Thumbs.db 158KB
Thumbs.db 157KB
Thumbs.db 156KB
Thumbs.db 150KB
Thumbs.db 147KB
Thumbs.db 141KB
Thumbs.db 106KB
Thumbs.db 18KB
Thumbs.db 16KB
Thumbs.db 14KB
Thumbs.db 13KB
Thumbs.db 12KB
Thumbs.db 12KB
Thumbs.db 9KB
Thumbs.db 8KB
Thumbs.db 8KB
Thumbs.db 8KB
Thumbs.db 8KB
028824.jpg 41KB
055077.jpg 41KB
028769.jpg 40KB
001345_hips.jpg 39KB
029410_rest.jpg 38KB
010995_rest.jpg 38KB
021495_folded.jpg 38KB
021580_folded.jpg 38KB
018810_rest.jpg 37KB
021820_folded.jpg 37KB
012194.jpg 37KB
016047.jpg 37KB
012263.jpg 37KB
012805_folded.jpg 37KB
000107.jpg 37KB
010075_rest.jpg 36KB
010075.jpg 36KB
000120_hips.jpg 36KB
025160_rest.jpg 36KB
010485_folded.jpg 35KB
010745_hips.jpg 35KB
001440_rest.jpg 35KB
005793_hips.jpg 35KB
001330.jpg 35KB
012805.jpg 35KB
006840_rest.jpg 35KB
026180_folded.jpg 35KB
035813.jpg 35KB
024840_folded.jpg 35KB
006885_rest.jpg 35KB
022420_folded.jpg 35KB
010485_rest.jpg 35KB
006820_rest.jpg 35KB
045950_rest.jpg 35KB
020940_folded.jpg 35KB
012303.jpg 35KB
022970_rest.jpg 35KB
055078.jpg 35KB
030955_rest.jpg 35KB
008435_hips.jpg 35KB
035830_hips.jpg 34KB
044605_rest.jpg 34KB
035915_hips.jpg 34KB
022895_folded.jpg 34KB
008795_hips.jpg 34KB
035515_hips.jpg 34KB
054685_hips.jpg 34KB
006985_hips.jpg 34KB
024926.jpg 34KB
028823.jpg 34KB
020397.jpg 34KB
054685_folded.jpg 34KB
040575_rest.jpg 34KB
028770.jpg 34KB
006100_hips.jpg 34KB
010170_folded.jpg 34KB
020956.jpg 33KB
008815.jpg 33KB
026120_folded.jpg 33KB
025500_folded.jpg 33KB
012195.jpg 33KB
013820.jpg 33KB
050145.jpg 33KB
039472.jpg 33KB
030900_rest.jpg 33KB
020918.jpg 33KB
012195.jpg 33KB
050160_hips.jpg 33KB
028306.jpg 32KB
031835.jpg 32KB
012324.jpg 32KB
051575.jpg 32KB
008670_hips.jpg 32KB
015662.jpg 32KB
055505_folded.jpg 32KB
039488.jpg 32KB
015895_rest.jpg 32KB
006800_rest.jpg 32KB
007680_rest.jpg 32KB
共 522 条
- 1
- 2
- 3
- 4
- 5
- 6
BryanDing
- 粉丝: 299
- 资源: 5583
下载权益
C知道特权
VIP文章
课程特权
开通VIP
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功
评论0