Project Name: Fast Depth Estimation
Most segmentation based algorithms are very computationally intensive which make them difficult to use for real time inference. This makes them less useful for both on-device as well as real time inference.
The goal of this project is to take the lessons from [Mobile AI Workshop 21 at CVPR and create a toolkit which would allow one to effectively train fast efficient models for segmentation tasks.
The process is really technically involved as it uses several optimizations, from Quantization to Knowledge Distillation.
Not having massive compute requirements not only helps in inference speed but is good for the environment and makes entirely new applications possible.
Examples of usage:
- Documentation on how to train a model using the toolkit
- Can be used with different data for Web3 AI Agents (chat bots etc) for Instagram like filters on images.
- Path planning for robots on-device
- Training Toolkit
- Example model for depth estimation from monocular camera (no special depth sensors) with fast inference capabilities
Tarun Kumar Vangani (me) - I am Research Engineer working on state-of-the-art ML models for 4+ years.
I would be excited to have more collaborators who can help push our system and hopefully put it to new unexpected usecases.