TargetV1: A Confidence-Gated, Open-Vocabulary Visual Recognition Pipeline Combining Self-Supervised, Zero-Shot, and Vision-Language Models

Modern open-vocabulary visual recognition typically relies on either a single large vision-language model or a fixed, hand-authored label vocabulary matched against image embeddings. Both approaches carry known weaknesses: single-model captioning systems produce fluent but frequently incomplete or biased descriptions of multi-subject scenes, while fixed-vocabulary classifiers cannot recognize anything outside their predefined label set and provide no mechanism for resolving low-confidence predictions. This paper presents TargetV1, a modular recognition pipeline that combines five publicly available models — a self-supervised vision transformer backbone (DINOv3), an open-set segmentation model (Segment Anything, SAM), a zero-shot image-text classifier (CLIP), a lightweight vision-language model (Moondream2), and an image captioning model (BLIP) — into a single system in which no component is asked to perform a task outside its area of demonstrated strength. The central contribution of this work is not a new model but a confidence-gated verification architecture: each detected region is assigned a calibrated confidence score by CLIP, and any score falling below an empirically determined threshold triggers an automatic, rate-limited external verification step against Wikipedia’s public image corpus, with the outcome — confirmed, inconclusive, or unconfirmed — reported transparently alongside a similarity score and source citation. The system additionally incorporates a lightweight, non-learned image-quality signal (Laplacian-variance blur detection) to flag when low confidence may be attributable to image degradation rather than genuine ambiguity. We document the complete engineering process required to bring such a pipeline to a working state on consumer Blackwell-architecture hardware, including a previously undocumented CUDA kernel incompatibility, a transformers library breaking change affecting custom model code, and several data-format mismatches between model outputs. We report qualitative results on a multi-subject test image containing over twenty distinct animal figures, discuss observed failure modes including a systematic underperformance of DINOv3 embeddings for cross-domain (photograph-to-illustration) similarity verification, and outline a path toward quantitative evaluation at scale.