Virtual meet is a video conferencing app where people can talk with each other. With this app, 3D model can be displayed in augmented reality. Gestures can be used to control different functionalities. Screen sharing can be done for presenting content towards other users. Participant can use the chat feature for communication.

• Mediasoup(mediasoup-client)
• Tensorflow.js(@tensorflow/tfjs)
• Ar.js(@ar-js-org/ar.js)
• Socket.io
• Mediapipe(@mediapipe)
• Material UI (@mui(

We use medaipipe hands library for web. The library was used to instantiate a pre-trained model. Webcam video frames are taken from the browser. Each frame obtained is passed to the model instance and a 21 point landmark is extracted from it. Gesture classification is done using the 21 point landmarks. On the basis of open and close of a finger’s status, gesture is classified.

SFU is implemented using MediaSoup library in nodejs. This nodejs media server allows us to create a multiparty video stream. For now SFU is only hosted locally due to limitation of resource and budget. Previously used heroku for deployment of peer-peer service does not support me- dia servers and other services available in internet for deployment are not viable for us, hence for now media server implementation is restricted to the local host

Threejs draws 3D environment on canvas. In order to get relevant 3D information from a video either a dedicated hardware are required such as in mobile devices or a mechanism needs to be in place to map 2D information into 3D. We used marker based AR to map 3D space from video. By capturing this 3D context from a monocular video stream, we are able to draw 3D objects relative to the video stream on the canvas. Now by making the canvas background transparent we are able to merge the canvas and video giving illusion of AR. This process can be further refined by programmatically merging the web cam video stream and canvas video stream. To achieve this we took the webcam video stream and painted each frame on a canvas, then layered the canvas stream on top of it. This new canvas was then used to produce a new stream with desired frames per second. By doing this we generate a single AR integrated video stream to view and transfer among peers. A dedicated button is used for turning AR on and off on the video stream. We have linked the toggling of AR mode to a ”Start AR” button so that user can turn on AR as required.

Loading of 3D model from Github There are various formats to create and distribute 3D models. Fbx, blend, mtl et are a few examples. However to transfer and load 3D models over the inter- net glTF/GLB are the standard 3D format. GLB is a binary form of glTF that includes textures instead of referencing them as external images. Thus providing comfort of a single source for transmission. Threejs has support for loading scenes and models from various formats. GLTFLoader provides simple methods to load glTF/GLB file formats from any URI source. 33 For this project’s demonstration, we pull freely availabe models from threejs repos- itories’ model examples. (https://github.com/mrdoob/three.js/tree/dev/examples/models/gltf)

User interaction with the AR model makes AR fun to use. Currently we listen to the keyboard buttons being pressed and when certain keys are pressed we perform respective mapped functions. The following functions are currently mapped to AR model • Rotation with 90 degree increment/decrement on X axis - ’a’ & ’d’ (Lower- cases) • Rotation with 90 degree increment/decrement on Y axis- ’w’ & ’s’ (Lower- cases) • Scaling by factor of 0.01 - ’W’ & ’S’ (Uppercases) • Reset Rotation - ’D’ (Uppercases) • Reset Scaling - ’A’ (Uppercases) We have also implemented a select input that lists the available models from threejs glTF model examples. When user selects a model, it is downloaded and replaces the current model on the screen. By default, a cube is loaded.