Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/ilirosmanaj/detect_kermit/llms.txt

Use this file to discover all available pages before exploring further.

A well-organised, representative dataset is the foundation of any reliable image classifier. For Kermit detection, the model must learn to distinguish frames containing Kermit the Frog from everything else — so both the positive (kermit) and negative (no-kermit) classes need enough variety to generalise beyond the training episodes. This guide walks through every step needed to get your image directories into the shape that ImageAI’s ModelTraining API expects.

​
Directory structure

ImageAI requires a strict folder layout. Create the following tree before running any helper scripts: 
data/
└── images/
    ├── train/
    │   ├── kermit/
    │   │   └── kermit-train-images/
    │   └── no-kermit/
    │       └── no-kermit-train-images/
    ├── test/
    │   ├── kermit/
    │   │   └── kermit-test-images/
    │   └── no-kermit/
    │       └── no-kermit-test-images/
    ├── models/
    └── json/
The models/ directory is where the trained .h5 file will be saved after training completes. The json/ directory holds model_class.json, which maps class indices to human-readable labels.

​
Data sources

Training data was extracted from Muppets episodes 1 and 2. Frames from these episodes are split into the kermit and no-kermit classes. Episode 3 is reserved as the test and validation set, providing a clean, held-out benchmark that the model never sees during training. The no-kermit class is deliberately diverse to prevent the model from learning superficial shortcuts. It includes images of:
  • Green frogs (visually similar to Kermit)
  • Human characters from the show
  • Nature scenes and backgrounds
  • Other Muppet characters
This variety forces the model to learn Kermit-specific features rather than simply detecting greenness or cartoon-like textures.

​
Step-by-step preparation

1

Extract frames from video

Use helpers/convert_vid2image.py to sample one JPEG frame per second from your source .avi video file. The script reads data/video.avi and writes numbered frames to data/videoframe/:
cd helpers
python convert_vid2image.py
See the Video to Frames page for a full breakdown of what this script does.
2

Organise frames into class folders

Manually review the extracted frames and move them into the appropriate subdirectories:
  • Frames containing Kermit → data/images/train/kermit/kermit-train-images/
  • All other frames → data/images/train/no-kermit/no-kermit-train-images/
Repeat this process for episode 3 frames, placing them into the corresponding test/ subdirectories.
3

Augment kermit images with rotation

Run helpers/rotate_images.py to triple the size of your kermit training set by generating rotated copies of every image:
cd helpers
python rotate_images.py
See the Rotate Images page for details on how rotations are applied and named.
4

Download supplemental images from Google (optional)

Use helpers/download_from_google.py to pull up to 100 additional Kermit images from Google Images into data/google-images/. Move relevant downloads into your kermit training folder afterwards:
cd helpers
python download_from_google.py
See the Download from Google page for configuration options.

​
Data enrichment via rotation

The rotate_images.py script reads every .jpg in data/images/train/kermit/kermit-train-images/ and produces three additional variants for each original:
VariantRotationFilename suffix
Right90°<name>right.jpg
Left−90°<name>left.jpg
180°180°<name>None.jpg
Because each original produces three new images, this augmentation quadruples the raw frame count (original + three rotations), effectively tripling the number of new samples. Exposing the model to Kermit at multiple orientations significantly improves robustness to camera angles and scene compositions it has not seen before.
Due to their large file sizes, the Muppets .avi video files are not checked into the repository. Only the helper scripts are version-controlled. You must supply your own video files and run the extraction scripts locally before training.

Build docs developers (and LLMs) love

Get started for freeTalk to us