A zero-shot object counting app is an app that counts objects without the need to specify what to count. Just
present an image, and the app automatically identifies visually similar objects in the image and counts
them.
This is enabled by an innovative architecture consisting of a ViT-based encoder and a convolutional decoder,
capable of extracting features from image and turning it into a "heat map" where each object is represented
by a dot.
Conventional methods are then used to count the dots in the "heat map".
Architecture
A ViT encoder followed by a convolutional decoder. The resulting number is the sum over decoder output.
Image resolution is 384 x 384.
Note
This app only provide a rough estimation of the object count. Estimations may be inaccurate or incorrect. Do
not rely on the results.
Please submit issues or feature requests here.
Privacy Policy
We do not collect any personal information. Your data is processed locally on your device.