-
Notifications
You must be signed in to change notification settings - Fork 0
Neural Net
Testing of various CV techniques.
About 50 to 60% inference accuracy upon initial testing using resnet-34 with 6 categories: plastic, metal, cardboard, paper, glass, and trash.
Use image segmentation to determine shape and surface area. Might be able to make depth calculations.
Based off Pytorch.
There are many neural networks architecture. Each implement different techniques and configurations. Generally, the more layers the network has, the more accurate it becomes, but it requires more computation power, such as GPUs. Once the model is trained, inference is done on either using cpu, gpu, or tpu (tensor processing unit).
Finding the best architecture for this project takes some trial and error. It needs to be lightweight, fast, low power, and accurate.
The first architecture I tested. It classifies images into glass, plastic, metal, paper, cardboard, and trash. Initial testing accuracy of 50-60% with an inference time of ~2 seconds.
A neural network designed to be deployed on devices with low computing power (cell phones, tablets, etc).
- Detectron2
- Inception v2
- DeepLabv3, semantic segmentation
- YOLOv3
Each identification technique has its own subdirectory and python virtual environment
- edge: contains py files for image segmentation
- fastai: contains py files for inference and classification