An image classifier and a language-translation system have different scopes. The inspected gesture route in LinguaAR follows a static hand-image classification pipeline.
This article describes the collaborative prototype's source. It does not claim authorship of every component or report a new model evaluation.
Isolate the detected hand#
The route uses a hand detector, takes the detected bounding region, and adds padding. The cropped image becomes the input for the classification preparation step.
The crop needs to preserve useful hand information without assuming the full camera frame already matches the model's expected input. Crop boundaries and absent detections are concrete cases for runtime evaluation.
Preserve shape within a fixed canvas#
The service resizes the crop according to its aspect ratio and places it on a white 300-by-300 canvas. A tall crop and a wide crop follow different resizing branches.
This supplies the classifier with a consistent input size while avoiding a simple stretch of every hand into a square. It does not establish robustness to every lighting condition, background, or hand position.
Interpret the output at its actual scope#
The Keras classifier chooses among 24 static labels. The inspected list omits J and Z. Selecting a label from one image does not demonstrate recognition of motion-dependent signs, continuous signing, or sentences.
Evaluation would need the model's actual dataset and test conditions before attaching an accuracy figure.
Verify the HTTP contract#
The inspected detection route expects an image field at /gesture/detect. The examined Flutter hand repository calls /gesture/hands with a file field. Those paths and field names need alignment or another verified route mapping.
A plausible processing pipeline is therefore only part of integration evidence. A real end-to-end check must connect the client request to the intended deployed handler and inspect the returned result.
The public implementation can be inspected in the Flask gesture routes and Flutter hand repository.