Collecting Data for Custom Object Detection
5 data collection techniques for training your custom detection model
I just read this awesome article from Towards Data Science, the author provided 5 different technique on collecting data.
- Publicly available open labelled datasets
We don’t need to label those images from scratch, we can save our time to fine-tune our model with this open labelled datasets
- Scraping the Web
I have tried this method last year, I downloaded 1000+ images from Google Images with a Python Script. You can search some scripts on GitHub on this topic.
- Taking photographs
If there are not enough target images on the Internet, you should probably take photos to create your own dataset, this is the most time-consuming method of these 5 method. And don’t forget it, once you have finished your photo-shooting journey you still have to LABEL these images😂😂😂
- Data Augmentation
It’s more complicated than Scraping the Web but I think it’s not that hard. We just need to adjust the photo with OpenCV or Pillow to Augmentation
- Data Generation
So far the most difficult technique you can use, write your own GAN( Generative Adversarial Nets) to create images from DeepLearning