Spatially-Aware Class-Agnostic Object Counting

2026-07-18

AI Summary

This work addresses the challenge of inaccurate counting of complex-structured objects in generic object counting, which stems from insufficient spatial modeling. To this end, the authors propose UpCount, a method that integrates multi-layer features from Vision Transformers (ViTs) and leverages Dense Prediction Transformers together with FeatUp to construct a spatially refined multi-scale pyramid. A proposal-and-verification counting head is further designed to generate high-fidelity density maps. By jointly exploiting multi-scale semantic information and fine-grained spatial details, UpCount significantly enhances the model’s perception of object structure and distribution. The method achieves state-of-the-art performance, reporting 12.39 MAE and 100.89 RMSE on the FSC-147 test set and 6.27 MAE and 8.79 RMSE on the CARPK vehicle counting benchmark.

Abstract

Generalised object counting aims to estimate the number of instances of an arbitrary object category from a single image, but many recent methods can struggle on structurally complex objects due to limited spatial modelling. We present \textit{UpCount}, a class-agnostic counter designed to better preserve spatial structure. UpCount strengthens the visual representation by extracting multi-layer features from a ViT-B/16 encoder and reassembling them into a refined multi-scale pyramid that is spatially refined using Dense Prediction Transformers and FeatUp, yielding features with improved structural and spatial sensitivity; a proposal--verification counting head then identifies repeated patterns and produces a density map for the final count. On FSC-147, UpCount achieves 12.39 MAE and 100.89 RMSE on the test set, and it transfers effectively to vehicle counting on CARPK (6.27 MAE, 8.79 RMSE). Code: https://github.com/r28112072-rgb/upcount

Problem

• object counting
• spatial modeling
• class-agnostic
• structural complexity
• instance estimation

Innovation

• spatially-aware
• class-agnostic counting
• multi-scale feature reassembly
• Dense Prediction Transformers
• FeatUp