Overview / Usage

We are annotating the indiancine.ma database which is a collection of Indian movies that are out of copyright. The annotated data could be used for searching and indexing the videos better. Also the false positives that might come from the model will also be an interesting way to assess how the model perceives history considering that the archive contains movies from 1950s' as well.

Methodology / Approach

Using an Single Shot MultiBox Detector (SSD) object detection approach, frames of videos can be run through the specified model and the related annotations can be saved onto a JSON file containing the annotations for each frame of the video. Scaling the model to be able to annotate multiple videos at the same time will require creating a server that is capable of Asynchronous requests.

Technologies Used

Tensorflow, Tornado, Python, OpenCV

Repository

https://github.com/Nikhil-Kasukurthi/Counting-people-video