autonomous-object-detection

Inspiration

We witness variety of environmental conditions when it comes to driving such as weather changes, dynamic changes in the surrounding environment, etc. Current detectors have been tested on data obtained from structured environments which are often not representative of real-world conditions. As a result of which, the need for data obtained from nonstandard sources is felt the most for data-driven algorithms to improve and test their generalizing capabilities. Autonomous navigation algorithms must perform well on multiple domains especially the ones with corner cases for safety purposes. Most importantly, we want to be able to learn from a large standard data distribution to efficiently learn features in an embedding space and learn progressively from domain-specific data without having access to earlier used data.

What it does

We address the problem of incremental learning and domain adaptation to some extent for object detectors to improve generalizing capabilities. Specifically, we tackle the problem of adapting from a standard data distribution to data obtained from the unstructured environment.

How I built it

Torchvision played a big role for with regards to object detection. I used multiple transfer learning approaches to tackle catastrophic interference during training step

Accomplishments that I'm proud of

• This work was accepted at ICCV 2019 workshop, this month.

What I learned

• Pytorch internals
• Torchvision's dynamics (from detection to core support)

What's next for autonomous-object-detection

• Improved support for 3D object detection models (FPNets, Lasernet)
• Dataset support (Lyft level 5, argoverse)

Built With

• python
• pytorch