Bootstrapping MLLM for Weakly‑Supervised Class‑Agnostic Object Counting

Core task: weakly-supervised class-agnostic object counting. The field organizes around several key dimensions. The first branch, Supervision Signal and Annotation Strategy, explores how to reduce labeling costs by moving from dense point annotations to weaker signals such as count-level supervision (Object Counts Labels[7], Count-level Supervision[18]) or even annotation-free methods (Annotation-Free Counting[43]). The second branch, Core Counting Approach and Architecture, encompasses the algorithmic backbone—ranging from density-map regression (TransCrowd[3], CrowdMLP[8]) to transformer-based frameworks (CountFormer[38], CountingDINO[37]) and segmentation-driven pipelines (Counting-by-Segmentation[16]). A third branch, Class-Agnostic Generalization and Multi-Class Handling, addresses the challenge of counting arbitrary object categories without retraining, often leveraging few-shot or zero-shot paradigms (Class-agnostic Few-shot Counting[9], Zero-shot Exemplars[33]). Application Domain and Task Specialization captures domain-specific adaptations (Apple Orchard Counting[4], Microorganism Enumeration[14]), while Surveys, Benchmarks, and Theoretical Frameworks (Class-Agnostic Counting Survey[12]) provide overarching perspectives and evaluation protocols. Recent work has intensified around two contrasting themes: reducing supervision overhead versus improving cross-category generalization. On one hand, methods like Unified Count-based Learning[39] and Temporal Classification Counting[41] push count-level supervision to handle multi-class scenarios with minimal annotation, while on the other, few-shot and zero-shot approaches (Few-shot Occlusion Counting[2], Training-free Baseline[47]) aim for rapid adaptation to novel object types. Bootstrapping MLLM Counting[0] sits within the count-level supervision cluster, closely aligned with Unified Count-based Learning[39] and Temporal Classification Counting[41], yet it distinguishes itself by leveraging multimodal large language models to bootstrap weak count signals into richer representations. Compared to Unified Count-based Learning[39], which unifies multiple count-based tasks under a single framework, Bootstrapping MLLM Counting[0] emphasizes the role of pretrained vision-language models in bridging the gap between minimal supervision and robust class-agnostic performance, reflecting a broader trend toward foundation-model-driven counting solutions.