A Quality Assurance Tool for Problem Formulation in Object Detection and Recognition

Keywords: problem formulation validation, domain similarity metrics, data audit, multimodal LLMs, hybrid systems, computer vision.

TL;DR: This paper proposes a three-level early-validation toolkit and a hybrid LLM-CV architecture to evaluate computer vision task formulations for logical and physical consistency before computationally expensive training begins.

Abstract:

The effectiveness of developing computer vision systems depends on the correctness of the initial task formulation and on assessing whether the target requirements are compatible with the capabilities of the chosen base models. This paper proposes a systematic approach to early task-formulation validation that identifies negative transfer risks, architectural constraints, and logical inconsistencies before computationally expensive design and training begin. The proposed approach is general and can be applied to problems solved using direct analytical methods, classical machine learning, and modern neural networks.

First, the paper organizes task-audit metrics into three levels: data analysis (KL divergence, MMD), model analysis (linear probing, Anchor Alignment Score), and training-dynamics analysis. The toolkit is shown to function as an effective Go/No-Go filter that evaluates not only whether fine-tuning or an algorithmic solution is feasible, but also whether the task statement itself is logically coherent and physically observable.

Next, the paper discusses a fundamental shift in recognition task formulation driven by multimodal large language models (MLLMs). It examines mechanisms for projecting visual features into a semantic space (MLP projectors, C-Abstractor) and new ways to specify tasks via natural-language instructions, which helps overcome the limitations of rigid class taxonomies. To achieve robustness in challenging conditions, a hybrid “Orchestrator–Executor” architecture is proposed: an LLM serves as the strategic node (semantic context validation), while specialized CV models (e.g., YOLO, SAM) provide tactical accuracy (geometric validation). Finally, end-to-end mission quality control metrics (Mission Success Rate) are introduced to link technical performance to business requirements.

Submission Number: 88