Our Repo consists of three major parts.
Each parts include their on README files with instructions, so follow the steps there.
Our project ran on a local machine. OS : Linux Mint 20 Cinnamon CPU : Ryzen 3900X GPU : RTX 3090 RAM : 64GB
Please note that some instructions might not work, especially those related to the training part, due to graphic card memory issue. In this case you have to modify batch size and learning rate to fit the memory.
OS should be Linux since our base library supports only Linux.
Crawler crawls images by keyword from internet.
Although making custom images is possible, it is recommended to use the crawler for convinience.
You could designate thye keyword you want to search for such as "brown wallet" and the crawler will crawl images for you.
Detailed information is included in https://github.com/paulkth2/fewshotLAF/AutoCrawler
Just having the images is not enough.
Since we've trained our base model with PascalVOC dataset, we've included a image labeler for the PascalVOC dataset with bounding boxes.
Detailed steps could be found on https://github.com/paulkth2/fewshotLAF/labelImg
The original library could be found below.
https://github.com/ucbdrive/few-shot-object-detection
We have modified parts of it to enable training with new custom objects.
Detailed steps to train with new objects are written in https://github.com/paulkth2/fewshotLAF/few-shot-object-detection
The full instruction regarding the library could be found on the original repo.
contributors: paulkth2, ellaide, sml0399, Tselmuun0624