Indoor Pet Detection Example
This project provides scripts necessary to train a model for dog detection in an indoor home environment, with Unity generated Synthetic Data, and evaluate your model. For this example, we will use synthetic data to train our model and real data from COCO and OIDSv6 to fine tune and test our model.
We also show how we can upload our own custom assets, and change randomizer parameters to change the data as required. Unity Computer Vision Datasets (UCVD) is the cloud platform used for generating large scale datasets.
Checkout our example at - Demo
Instructions
- Prerequisites
- Create a synthetic dataset for dog detection with UCVD
- Training, Fine-tuning & Evaluation
- Results
Learning More About Unity Synthetic Data
- Unity Computer Vision Datasets
- Overview on how UCVD works
- Unity Perception
- Create your own assets on UCVD
Sample Datasets
Results
Pre-training | Fine tune(real data) | AP | Data generation time(mins) |
ImageNet + no synthetic | 1200 | 51.16 | --- |
ImageNet + 5k synthetic | 1200 | 57.18 | 33 |
ImageNet + 10k synthetic | 1200 | 58.42 | 35 |
ImageNet + 40k synthetic | 1200 | 63.03 | 50 |
ImageNet + 100k synthetic | 1200 | 65.4 | 84 |
Scratch + no synthetic | 1200 | 23.16 | --- |
Scratch + 5k synthetic | 1200 | 36.16 | 33 |
Scratch + 10k synthetic | 1200 | 36.97 | 35 |
Scratch + 40k synthetic | 1200 | 47.3 | 50 |
Scratch + 100k synthetic | 1200 | 53.78 | 84 |