MambaCount: Efficient Text-guided Open-vocabulary Object Counting with Spatial Sparse State Space Duality Block

Researchers have proposed new frameworks for text-guided open-vocabulary object counting, a challenging task in computer vision. MambaCount, RT-Counter, and Robust-TOOC are among the latest developments in this field.

MambaCount, introduced in a paper on arXiv^[1], is an efficient framework built on the Spatial Sparse State Space Duality (S^4D) block. It alleviates the dependency constraints introduced by causal modeling in Mamba and reduces the unconstrained high entropy in spatial token responses. MambaCount achieves state-of-the-art performance among methods without secondary querying on the FSC-147 dataset. Meanwhile, RT-Counter, proposed in another arXiv paper^[2], uses a Visual Prototype Textualization module and Weaving Transformer layers to achieve a competitive MAE of 13.30 on FSC147 while operating at 112.48 FPS^[2]. RT-Counter is also 7.4x faster and over 4x more parameter-efficient than existing leading methods. Robust-TOOC, introduced in a third arXiv paper^[3], is a benchmark for evaluating Text-guided Open-vocabulary Object Counting under diverse corruption conditions, covering six representative degradation types. The Dual-TTT method, proposed alongside Robust-TOOC, updates only the Text-guided Lightweight Denoising module (TL-Denoiser) during test-time training and is annotation-free.