/detect

Zero-shot object detection in the browser with CLIP and TensorFlow.js

Primary LanguageJavaScript

Zero-shot object detection with CLIP and TensorFlow.js

The app loads a TensorFlow.js CLIP model from modelzoo.js and h5 weights from HF. CLIP encodes images and texts into a 512-dimensional embedding space. The app captures the webcam stream and encodes it into a 512-dimensional vector. Then it compares the vector with the embeddings of provided classes and if the target class is found, sound is played and the class is displayed.