🎉 Approved by NVIDIA!

This is an AI that can determine whether or not are you touching grass. My inspiration when creating this project was a popular [meme](https://knowyourmeme.com/memes/touch-grass) from the internet, where people jokingly tell each other to "go outside and touch grass" when they’re spending too much time on the computer.

Inference

Clone the repository:

git clone https://github.com/marks0kolov/touch_grass.git
cd touch_grass

Install the required packages

pip3 install -r scripts/requirements.txt

Run inference on an image with the provided python script:

python3 scripts/check_touching_grass.py path_to_image.jpg models/grassNet/model.onnx models/poseNet/model.onnx

(don't forget to replace path_to_image with the actual path to image)

The script will load both ONNX models, use both of them to check if the image contains grass and a hand and print the result. To check for only grass or hands run scripts/check_grass.py and scripts/check_hand.py accordingly.

GUI

The AI also has a GUI with a timer that can remind you to touch grass from time to time and check if you've really done it.
You can set a timer:

And then when it finishes the program will ask you to upload a photo of you touching grass:

The AI will then recognize whether or not you uploaded a real photo and not stop the alarm until you uploaded a real photo of you touching grass

After that the timer setting page will appear again.
To use the interface, you need to run a local server from the root of the project with

python3 -m http.server 1008

and then navigate to http://localhost:1008/gui

Training

Firstly, I took 10.5k images of grass from various datasets on the internet (more info in resources). Then i added an image of a hand to 80% of those images, so that the model also accepts images with hands covering some part of it. I also downloaded 10k of indoor images from the Places365 dataset.
Secondly, using the train.py script from the jetson-inference GitHub repo, I fine‑tuned an Image ResNet‑18 on this dataset for 40 epochs. The best model with an accuracy of 99.82% and a loss of ~0.009 on the validation images was the checkpoint from epoch 36, so i took it as the main model.

I also took the original Pose ResNet-18 Hand ONNX model from the jetson-inference repository to detect hands on images.

Resources

• Datasets

  • Deep Learning. 2025. "Capstone Project – Grass Species." Roboflow Universe. Roboflow, Inc. https://universe.roboflow.com/deep-learning-4rbtb/capstone-project-grass-species
  • Iowa State University. 2024. "Grass O0vum v1." Roboflow Universe. Roboflow, Inc. https://universe.roboflow.com/iowa-state-university-krhld/grass-o0vum
  • Usharengaraju. n.d. "GrassClover Dataset." Kaggle. Accessed July 24, 2025. https://www.kaggle.com/datasets/usharengaraju/grassclover-dataset
  • Timofeymoiseev. n.d. "Grass Detection Dataset." Kaggle. Accessed July 24, 2025. https://www.kaggle.com/datasets/timofeymoiseev/grass
  • Jonasdahlqvist. n.d. "Grass‑NoGrass Dataset." Kaggle. Accessed July 24, 2025. https://www.kaggle.com/datasets/jonasdahlqvist/grass-nograss-dataset
  • CSAILVision. 2017. "Places365‑CNNs for Scene Classification". GitHub. https://github.com/CSAILVision/places365. Accessed July 24, 2025.

• Models

  • Image ResNet-18
  • Pose ResNet-18

• NVIDIA Jetson Orin Nano

• Dusty-nv. Updated October, 2024. "jetson-inference", GitHub. https://github.com/dusty-nv/jetson-inference.