Hey there. I struggled to find an interesting research topic for my Master thesis so I can put my full effort and finish it on time. But it appeared that this is the first challenge that takes so much time. For now, I have 2-3 topics that make me curious think about them, and I am mostly focused on "Azerbaijani Sing Language to text" which is a project that will be used by the Government Services of Azerbaijan (in short ASAN) which serve to the citizens of Azerbaijan.
The problem is that it is very hard for the deaf community of Azerbaijan to communicate (and clarify their issue) in daily life, especially in ASAN branches. In order to ease their life and speed up the service process someone is needed to build the API which receives a video of a person signing as an input and outputs the result as a sentence in Azerbaijani. In addition to that, for a full conversation between a deaf person and a serviceman API should be able to read a text in Azerbaijani and create a video (separate image frames combined together) which describes all the signs for the sentence.
In order to build and train the model, I will need to gather a lot of images containing signs that are used in Azerbaijani Sign Language with a correct label. And this is the tricky part as there is no source for any image containing Azerbaijani Sing Language, and even if I find some on social networks there is no label mostly.
So, I came up with some possible solutions to collect some data with correct labels.
First, a Telegram bot. I will build and deploy a bot in Telegram as it is commonly used among people nowadays in Azerbaijan. I will try to share that app through all the channels, including ADA university, ASAN community (as they have thousands of volunteers), social networks, etc. The bot will ask a user to upload a video and then send a sentence which contains words corresponding to each image in the video.
Second, TV shows. I do remember that there were a few shows in the Azerbaijani Television channel that were supporting the deaf community by translating everything to sing language and showing it on the bottom right corner of the video. I will try to reach them, persuade them that with their help (and data) I will be able to come up with a solution. I hope this will be a great source of images for formal language of Azerbaijan.
Third, surprise support from Ministry of Social Affairs of Azerbaijan. As this project is fully supported by Ministry of Social Affairs of Azerbaijan we have a ML scientist working on this project in the Ministry. I will contact her and check whether she has more resources for data collection.
As I think about this project, I believe that there will be many emerging resources through the process of data collection and analysis.
This is a great post. It sounds like you have some strong chances to get access to data. Do you have a guess of how much data you need? I think there are interesting questions about exactly what kind of data you can get --- lots of data in good conditions from a few people? a little bit of data from many people? Some of both?
All of these things change what kind of systems you might be able to train!