I had an idea to implement a real-time database demonstration. But unfortunately, I am not able to get a relevant dataset.
So, I asked my university childhood friend to suggest an idea. He gave an innovative solution for detecting the right books in the library. Also, he helped me with the idea formation and dataset collection process.
He sent images of multidisciplinary books that he took from the library. He is a reader, so he felt difficulties finding the right book in a big library base. In my opinion, the human eye could make some mistakes but not the AI-powered cameras backed by deep learning and object detection.
He took 83 pictures of Books from his OnePlus Nord phone in different angles, shuffled, one book above another, upside down, blurred, etcetera. Thanks for your precious time!
Then, I have decided to Label the pictures using LabelImg tool. I wish, could work well.
The first image is a sample Annotation Screenshot Pascal VOC format and the second one is the labelled names of corresponding books
Now, I have to load the images to Detectron 2
Thanks to Roboflow – They’re offering free images to upload in our account and helps to feed them into Computer Vision algorithms like Detectron 2, Yolo, etc. via a link
1. You have to create an account
2. Upload the images
3. We need to split the data into training, validation, and testing as per your wish – see the guidelines
4. And finally, Generate the dataset archive in COCO format – Copy the link for your dataset
Note: Don’t share the link unnecessarily
5. Paste the copied link in between the code
6. Train Detectron 2 sequentially as given in the GitHub link-> https://github.com/freefacts/Detectron-2_Books Just Run the code without changing anything
You could change the backbone architecture, Iterations to train the model and the Number of classes
OMG! It took only 1Hr 15mins to train Detectron 2 with Faster R-CNNThe results will shock you, Actually – I was surprised
Booyah! Almost 99% accuracy on tested imagesAmazing right🔥
Finally, the above slideshow represents the tested images on different angles and styles. In the end, the result is very much like, as we expected.