https://github.com/mayukhdeb/deep-chicken-terminator
tracking wildlife in minecraft using deep learning :hocho: :chicken:
https://github.com/mayukhdeb/deep-chicken-terminator
computer-vision data-augmentation deep-learning heatmaps minecraft object-detection object-tracking opencv pyautogui pytorch
Last synced: 7 months ago
JSON representation
tracking wildlife in minecraft using deep learning :hocho: :chicken:
- Host: GitHub
- URL: https://github.com/mayukhdeb/deep-chicken-terminator
- Owner: Mayukhdeb
- Created: 2020-03-01T13:06:09.000Z (over 5 years ago)
- Default Branch: master
- Last Pushed: 2020-04-10T05:27:06.000Z (over 5 years ago)
- Last Synced: 2025-02-28T10:34:00.685Z (8 months ago)
- Topics: computer-vision, data-augmentation, deep-learning, heatmaps, minecraft, object-detection, object-tracking, opencv, pyautogui, pytorch
- Language: Jupyter Notebook
- Homepage:
- Size: 17.9 MB
- Stars: 8
- Watchers: 2
- Forks: 0
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# deep-chicken-terminator
deep learning to track (and possibly kill) chickens in minecraft :hocho: :chicken:
[](https://nbviewer.jupyter.org/github/Mayukhdeb/deep-chicken-terminator/tree/master/)
.
## Step 1 - collecting training data for the deep neural network
.
* This was mostly just me taking cropped screenshots of animals while playing the game, took about 40 screenshots of each animal.
* The 20 screenshots of each animal were then augmented and got 500 samples of each.
* The dataset is very small, but it works anyways for now
## Step 2 - training a deep learning model on the samples
.
* The architecture was kept intentionally small so that it keeps a good response time on the live feed
* The dataset had only 2000 images sized at 50*50, so training barely took any time
## Step 3 - Collecting more training data with the trained model
.
* this is done by saving the frames which give a probaility of over 99% on the pre trained model
* these saved images are again used for further training, which means `hunter()` is getting better and better.
## Step 4 - detecting a and tracking chicken (or any animal for that matter) with the mouse using the trained model
.
.
* This was done using a custom ```detect_animal_numpy()``` function which iterates through the image with a certain kernel size and a stride size, and feeds each sample to the trained NN (nicknamed hunter)
* A heatmap is then generated from the output of the NN which gives us a probability distribution over the image of a certain animal ( chicken, pig, or panda)
* why use heatmaps instead of rectangles ? because they look cooler.
.
## Step 5 (and probably the final step) - Train `hunter()` to detect fellow villagers and wipe out whole villages
* Chickens are just an excuse, it can be easily modified to shoot arrows on anything that moves.
* This is yet to be done.