- The results of the assignments are limited by the quality and quantity of data available to the applicant. For instance, only the data available in the JSON format after retrieving a certain task is used for the QC.
- If more data were available in the aforementioned JSON, then potentially better QC is always possible. For instance, if
image_metadata
attributes such asdate_time
or eventags
, were available, then images taken after sunset could be rejected due to poorer quality of annotations.
- In a virtual environment, install the following:-
pip install requests
pip install json
pip install os
pip install python3
pip3 install jupyter
pip install -U jupyter
- Start the notebook server from the command line:-
jupyter notebook
You should see the notebook open in your browser.
DONE
Retrieved tasks using project name Traffic Sign Detection
and task status Completed
API used : https://api.scale.com/v1/
Authentication used - Bearer Token - base64 form of my_api_key
DONE
To maintain consistency in labelling, annotation rules states background_color
as not_applicable
should be used only for the non_visible_face
label.
The implemented QC makes sure that this convention is followed.
DONE
-
Largest signs are well annotated. i.e compare an annotation to the image resolution. Pick the top 5 largest annotations, and make sure they have enough information.
-
High Priority signs are well annotated.
- e.g. traffic lights -> non zero
truncation
/occlusion
ortraffic_light
construction_sign
- e.g. traffic lights -> non zero
-
No low-light photos. check
timeofday
tag -
No photos at curves or roundabouts due to obscurity and poor angles
-
Increasing level of severity based on the label. E.g. if error in traffic_signs, then higher priority than for information or policy signs.
Typical Workflow of ML project
source
/ generate
--> annotate
--> manage
/ evaluate
--> automate
- Biggest impact of QC is at
source
. - Next best option is at
annotate
stage. - 3rd best option is at
evaluate
and have a feedback loop togenerate
.