A repo of my exploration of open-cv, started with the idea of being able to detect when my cat is on webcam, spiralled from there.
OpenCV2 includes an out-of-the box cat classifier(s): haarcascade_frontalcatface.xml and
haarcascade_frontalcatface_entended.xml. These are fairly decent at getting most cat faces, but for some reason
struggled with my cat, Hammy.
Python3 installation with opencv (v4.x) pip install opencv-python
So, basically for training we need a decent sized set of positive images (containing the thing you are trying to detect), as well as a set of negative images of similar composition, but not containing your target object. At least 100 of each is a good benchmark.
In root there is a folder for negative images (negative_images) and one for positive images
(positive_images), you can either drop your own existing photos into these folders, or you can use the capture_webcam
function which will add a timelapse series of images from your webcam to, the raw_images folder, from here you can run
categorise.py and each photo will appear and you can press p for a positive image, n for a negative image, i for
an inconclusive image, and these will be moved to the appropriate folders accordingly.
From here there are a few more steps. You need to make summary files of the negative and positive images you have captured.
Starting with the negatives, simply run generate_negative_txt.py and a file called neg.txt will be generated.
The positive images are a bit more involved, and this is where our opencv v3 docker image comes into play. The steps for this are:
- Pull this handy pre-made docker image (V3)
docker pull jjanzic/docker-python3-opencv:contrib-opencv-3.4.2 - Build through docker desktop - including mounting the project directory, so you can access files between your local repo and your docker instance.
- Run you v3 container:
docker exec -it opencv3 /bin/bash
From within the bash terminal of your docker image you can then run the scripts necessary to generate the summary of the positive images.
Run this to annotate your images - basically click once to start a box, drag, and click again to expand and finish that
box, follow the instructions printed in the terminal to see how you accept changes and move on to the next image.
opencv_annotation --annotations=pos.txt --images=positive_images/. This should result in a file called pos.txt being
generated in root.
This command will then generate a vector file needed to train the HAAR cascade model.
opencv_createsamples -info pos.txt -w 24 -h 24 -num 1000 -vec pos.vec
Finally, to train your model:
opencv_traincascade -data cascade -vec pos.vec -bg neg.txt -w 24 -h 24 -numPos 130 -numNeg 100.
NOTE: numPos needs to be a bit smaller than the number of positive samples. I think numNeg can just be whatever.
Then you should have a nice model generated in cascade called cascade.xml.
Running main.py will implement this on your live webcam feed.