325 karma · joined February 26, 2015
PhD student at UMD College Park mendeza@umd.edu
I agree with monkeynotes that you have to be mindful on how much time you spend collecting bookmarks, I felt consumed at one point and unproductive because you skim, you save, and you find other links to the original link you are saving. Probably set a time limit to how long you spend.
Slides from Josh Tobin is a great introduction: http://josh-tobin.com/assets/pdf/randomization_and_the_reali...
http://josh-tobin.com/assets/pdf/BeyondDomainRandomization_T...
And a really cool project implementing synthetic generation of text in images: https://github.com/ankush-me/SynthText
My guess is Apple trained on their own massive dataset and better architectures that make their systems better than off the shelf ocr.
I am working in this area and one way I see good improvement is training an object detector to first detect words in an image, then you can pass that through tesseract/OCR software. Besides that, finetuning tesseract on data you want strong performance would be the next best alternative.
Here is a cool article from dropbox how they engineering their OCR system: https://blogs.dropbox.com/tech/2017/04/creating-a-modern-ocr...
This link below is a really cool project that showcases what building your own OCR engine would look like!
https://github.com/awslabs/handwritten-text-recognition-for-...
Template Matching scans an image with your template, smaller image and finds the location with highest cross correlation: https://opencv-python-tutroals.readthedocs.io/en/latest/py_t...
Template matching assumes the example image wont deform or take on different orientations.
Popular Feature Extraction finds and derive more robust features that help find objects in the scenarios of distortion, orientation, and scale.
Link for feature detection and matching: https://opencv-python-tutroals.readthedocs.io/en/latest/py_t...
Pyimagesearch does a cool method where they enable template matching to work at different scales. It ignores color information (by using edge detection), which could be a critical feature if the object has specific color:
https://www.pyimagesearch.com/2015/01/26/multi-scale-templat...
Also, I would recommend this repo to better understand how object detection models are evaluated as well: https://github.com/rafaelpadilla/Object-Detection-Metrics
Also the TF Object Detection API has a suite of pretrained SOTA object detection models https://github.com/tensorflow/models/tree/master/research/ob...
What makes object detection really valuable is being able to finetune a model to detect objects that are relevant to your problem. However collecting, labeling, takes a lot of time since the best way is to manually label. But 30-50 hrs of manual labeling can get you some great results when you are leveraging pretrained models. Building a tool to manage data labeling and training for finetuning object detection models would be very valuable.
This is insightful, but how does one deal with determing if your dataset is non-random? I can imagine manually inspecting the data and using cross-validation are ways to help identify skewed datasets. Are there any other ways to test if your data is non-random?
Are you applying for junior dev roles? Maybe finding a career counselor of some sort can help. Reach out to your friends and ask for their honest advice of what they think you should do/add. Leverage your personal connections to get your first first full time job!
Ultimately, you have to persevere and be scrappy to find your first job. IMO the hardest part in your career is when you first enter the workforce. Always try to stand out and market yourself on how you would be a huge asset to the employer.
Last thought, if you are looking to join a great school/community of aspiring junior engineers and web developers, apply to Lambda School. I recommended my own brother and friends, and I have seen huge growth in their dev skills and abilities. I hope this helps!