This repo contains the server-side source code for Captivate!, a system that analyzes real-time video and audio streams to provide language guidance for parents in play situations.
The system separately processes video and audio streams, with the goal of extracting attention-related cues and combining them to infer a joint attention distribution. The joint attention distribution provides the basis for the context-relevant phrases that are recommended to the parent in real-time. Below is an overview of the system.
The processing pipline for video and audio is in the modelserver directory.
The main process is modelserver.modelserver.run_server.py, wihch starts
several parallel processes that fetch and handle real-time data streams from an
RTMP video server. Each process writes model inference outputs to a shared
queue in the main process, which combines the information to calculate the
joint attention distribution.
In our study, we used three cameras, so therefore launch three visual processes that perform the same pipeline on each of the video streams. Note that this number can be changed easily.
Finally, communication between the server and client is done in real-time with WebSockets.
Currently the code serves mainly as a reference, as running takes quite a bit of setup (and the client code is not open yet). That said, the Docker image for the modelserver can be built and run with.
./build.sh
./run.shWe would like to thank and mention the following repositories and their authors for providing the models used in this system.
- Gaze following: https://github.com/ejcgt/attention-target-detection
- Face detection: https://github.com/deepinsight/insightface
- Object detection: https://github.com/facebookresearch/detectron2
- Korean morpheme analysis: https://github.com/kakao/khaiii