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.

kermit_model_evaluation.py loads the trained ResNet model and runs inference via ImageAI’s CustomImagePrediction class. It accepts either a single image, a comma-separated list of images, or a video file as input, and returns class probabilities for both kermit and no-kermit on each frame or image. For video input, the script also burns the prediction text directly onto each extracted frame and saves the annotated images to disk.

​
How the evaluation script works

On startup, the script:
  1. Instantiates a CustomImagePrediction model and sets it to ResNet architecture.
  2. Loads the trained weights from data/images/models/kermit_finder.h5.
  3. Loads the class label map from data/images/json/model_class.json.
  4. Accepts the -t flag (image or video) to select the input type.
  5. Accepts the -f flag with the path (or comma-separated paths) to the target file(s).
  6. Returns probabilities for both kermit and no-kermit classes for every input.

​
Predicting a single image

Pass -t image and the path to your image file: 
python kermit_model_evaluation.py -t image -f kermit.jpeg
Expected output: 
Predicting the kermit.jpeg image
 kermit: 99.87 no-kermit: 0.13

​
Predicting multiple images

Supply a comma-separated list of file paths to the -f flag — no spaces around the commas: 
python kermit_model_evaluation.py -t image -f image1.jpg,image2.jpg,image3.jpg
The script iterates through each path in order and prints the prediction result for each file in turn.

​
Predicting a video

Pass -t video and the path to an .avi file: 
python kermit_model_evaluation.py -t video -f MuppetsEpisode3.avi
The script processes the video as follows:
  • Frame extraction — OpenCV reads one frame per second (at 1 000 ms intervals) for the entire duration of the video.
  • Frame storage — Each extracted frame is saved as a JPEG to episode3_results/ep3_frameN.jpg, where N is the frame index.
  • Async batch prediction — All frames are dispatched concurrently using Python’s asyncio event loop via a gather_dict utility, making full use of available compute without waiting for one frame to finish before starting the next.
  • Annotation — After all predictions return, OpenCV writes a text banner onto each saved frame image showing the kermit and no-kermit probabilities (e.g. kermit 99.87% no-kermit 0.13%), and overwrites the JPEG on disk.

​
The predict_image async function

The core prediction primitive is an async function that wraps ImageAI’s synchronous predictImage call: 
async def predict_image(image_name: Union[str, np.ndarray], model: CustomImagePrediction) -> dict:
    """Predicts a given image with the supplied prediction model"""
It accepts either a file path string or a NumPy array (allowing it to be called directly on OpenCV frame data) alongside the loaded CustomImagePrediction model instance. The return value is a dict mapping each class name to its probability formatted as a two-decimal percentage string — for example: 
{'kermit': '99.87%', 'no-kermit': '0.13%'}
This dictionary is what gets serialised into the text banner drawn on video frames, and printed to stdout for image inputs.

​
Known limitations

  • Kermit-like false positives — The model may occasionally misclassify visually similar characters (such as green frogs or other amphibian-like Muppets) as Kermit. This is noted in the project README as a known behaviour, and arises because the decision boundary between “green frog” and “Kermit” is inherently subtle. Adding more diverse no-kermit examples — especially of green frogs — during training can reduce this error.
  • Video format — The video prediction path is designed around .avi files and uses cv2.VideoCapture directly. Other container formats may work depending on your OpenCV build, but only .avi has been tested.
  • Output directory — All annotated video frames are written to a flat episode3_results/ directory relative to wherever you run the script. There is currently no option to change this output path via CLI flags.
For a full list of available command-line flags (including --file_type / -t and --files / -f), see the CLI Reference.

Build docs developers (and LLMs) love

Get started for freeTalk to us