We are a small company with a couple of successful products around car identification. See https://platerecognizer.com/ for details. We are looking for someone with a good understanding of how to build models from scratch and acute attention to detail. Your first project will be an OCR reader (see below). Contact me if you are interested.
We are working on a shipping container ID reader. It will be composed of an object detector and an OCR module. The detector will output bounding boxes. The OCR will take as input the cropped images based on those bounding boxes. Your job is to work on the OCR only. Here are the requirements:
- Decode ID + ISO type
- It should work on both horizontal and vertical text. It’s ok to use 2 separate models.
- Use Tensorflow and Python. Deliverable must include training + inference code.
- The code has to be based on this architecture : CNN + RNN + CTC loss. Like this project OCR model for reading Captchas
- We are providing the annotated training data. Around 800 images. Each has 3 pieces of text.
- Inference speed: 100ms/image on a 5000 Passmark CPU.
- Support all format variations at 95% accuracy. We will test using images from the same distribution but not the same images as the training data.
Here’s what the dataset looks like: