Difference between revisions of "Data Annotation"

From CSWiki
Jump to: navigation, search
 
(11 intermediate revisions by 2 users not shown)
Line 1: Line 1:
* Bounding box of ImageNet images
+
==Bounding box Label==
  Annotation file format:  
+
* Annotation file format:  
 
   Img_W Img_H xmin xmax ymin ymax
 
   Img_W Img_H xmin xmax ymin ymax
 +
* What we have collected:
 +
  router
 +
  guitar
 +
  swivel chair
 +
  electric organ
 +
  vibes
 +
* What we plan to collect:
 +
  "device" and "equipment" of instrumentality (instrumentation), like computer, mouse, stapler, ...
 +
 
 
* Problems of Current Approach
 
* Problems of Current Approach
# some images are too large
+
# some images are too large to be displayed within the screen, so have bad annotation.
 
# the "null" string in result should be replaced by -1 -1 -1 -1
 
# the "null" string in result should be replaced by -1 -1 -1 -1
# the following plot is that, if each image has a certain number of annotations (x axis),  
+
# the following plot is that, if each image has a certain number of annotations (x axis), how many images can be inferred a true bounding box.
how many images can be inferred a true bounding box.
 
  
[[image:valid_img_ratio_with_respect_to_num_of_annotation.png|400px]]
+
[[image:valid_img_ratio_with_respect_to_num_of_annotation.png|800px]]
 +
Attach:valid_img_ratio_with_respect_to_num_of_annotation.png
  
 
By the plot, we suggest using 6 annotations for each image to label.
 
By the plot, we suggest using 6 annotations for each image to label.
 +
 +
*Change random display of images to sequentially presenting a list of images to workers. Store a local counter in a text file or database.
 +
-Need to consider keep using text files or switching to database
 +
 +
==View Point Label==
 +
* Ask labelers to do a selection with each branch to be a thumbnail of a frame sampled from video sequences.
 +
 +
We plan to sample those frames with equally distributed angles.

Latest revision as of 19:42, 5 January 2009

Bounding box Label

  • Annotation file format:
 Img_W Img_H xmin xmax ymin ymax
  • What we have collected:
 router
 guitar
 swivel chair
 electric organ
 vibes
  • What we plan to collect:
 "device" and "equipment" of instrumentality (instrumentation), like computer, mouse, stapler, ...
 
  • Problems of Current Approach
  1. some images are too large to be displayed within the screen, so have bad annotation.
  2. the "null" string in result should be replaced by -1 -1 -1 -1
  3. the following plot is that, if each image has a certain number of annotations (x axis), how many images can be inferred a true bounding box.

Valid img ratio with respect to num of annotation.png Attach:valid_img_ratio_with_respect_to_num_of_annotation.png

By the plot, we suggest using 6 annotations for each image to label.

  • Change random display of images to sequentially presenting a list of images to workers. Store a local counter in a text file or database.

-Need to consider keep using text files or switching to database

View Point Label

  • Ask labelers to do a selection with each branch to be a thumbnail of a frame sampled from video sequences.

We plan to sample those frames with equally distributed angles.