Computer Vision for Objective Medication Adherence Assessment: From Pill Recognition to Ingestion Verification and Clinical Benefit, with Comparative Evidence from Non-Visual Digital Adherence Technologies

Medication nonadherence remains common despite decades of measurement research, and no widely adopted method documents, dose by dose, that a medicine was actually swallowed. Computer vision now touches three linked tasks: image-based pill identification, automated verification of ingestion from smartphone or wearable video, and camera-supported supervision in tuberculosis programs and trials. This narrative review examines that literature critically, and treats pill-exit monitors, electronic bottle caps, ingestible sensors, and SMS supervision as comparative evidence, because outcomes have been measured for those technologies. Deep learning performs well on curated pill benchmarks, but the main public reference resource was retired in 2021, consumer-facing identifiers show uneven accuracy, and multimodal models succeed mainly where packaging text is legible, which sidesteps the unpackaged tablet. Ingestion verification can in principle displace weaker proxies. Several large randomized evaluations of digital adherence technologies nonetheless report better measured adherence without better clinical outcomes, and the leading ingestible sensor company entered bankruptcy despite approval: a commercial event, not a verdict on clinical effect. Deployment raises unresolved questions about privacy, surveillance of vulnerable groups, regulation, cost, equitable access, and the readiness of clinicians and pharmacists to act on a verified-dose record. Accuracy alone no longer settles what a system must show, since evidence of clinical utility has become an equally demanding requirement, while domain shift, worn imprints, low-shot class structure, consumer-image variability, and confident misidentification remain unresolved. We outline the benchmarks, trials, subgroup reporting, and governance needed to test that proposition, for which evidence is thin.