Create a handler for people based images

[rohanakut]

This issue mainly deals with creation of handlers for people category. The additional ML models implemented in #455 would be used in creating renderings for these images. An example image for which this handler should work is mentioned below

(Source - https://www.theguardian.com/uk-news/2022/aug/01/david-trimble-funeral-politicians-unite-pay-tribute)

The following should be the tasks of the handler

• [x] Determine the category of the image. In this case the category should be people
• [x] (Updated on Aug 1) Since the Azure graphic tagger is not reliable, we need to think of a custom check for the same.
• [x] Remove low confidence objects and very small objects
• [x] Check if certain people occupy bigger area than others. The rendering strategy should change depending on the situation
• [x] Run the emotion recognition
• [ ] Add position information
• [ ] Add information related to clothing(link and link)

There are a few additional models which could be run. However, these models need to be tested before implementing them

• [x] Celebrity Recognition: It would be helpful to determine if the photo contains a celebrity. However, the accuracy of the model needs to be tested, especially the false positive rate for this model
• [ ] Including information from the title: The initial plan was to use NER model to use the nearby context to get better renderings. However, there are several problems with NER implementation. Hence the implementation of NEr is still under question.

[rohanakut]

The azure classifier does not classify image as "people", despite the image having multiple people. For example the following image the azure classifier returns 400 error. (Source - https://www.cbc.ca/news/canada/first-person-colonization-my-family-1.6515342)

Since the Azure Classifier is not reliable for detecting "people" category, there needs to be a custom check to determine of there are people in the image and if the image belongs to people category. The same has been added to task list.

[rohanakut]

Update: The first renderings were generated using the tasks described above. However, the renderings could contain more information like celebrity names, clothing information etc. Hence I have updated the task list to add these preprocessors.

[rohanakut]

@jeffbl closing this, since it is completed