A Low-Shot Object Counting Network With Iterative Prototype Adaptation

We consider low-shot counting of arbitrary semantic categories in the imageusing only few annotated exemplars (few-shot) or no exemplars (no-shot). Thestandard few-shot pipeline follows extraction of appearance queries fromexemplars and matching them with image features to infer the object counts.Existing methods extract queries by feature pooling which neglects the shapeinformation (e.g., size and aspect) and leads to a reduced object localizationaccuracy and count estimates. We propose a Low-shot Object Counting networkwith iterative prototype Adaptation (LOCA). Our main contribution is the newobject prototype extraction module, which iteratively fuses the exemplar shapeand appearance information with image features. The module is easily adapted tozero-shot scenarios, enabling LOCA to cover the entire spectrum of low-shotcounting problems. LOCA outperforms all recent state-of-the-art methods onFSC147 benchmark by 20-30% in RMSE on one-shot and few-shot and achievesstate-of-the-art on zero-shot scenarios, while demonstrating bettergeneralization capabilities.