Indian small and medium manufacturers operate a vast installed base of legacy CCTV infrastructure — Hikvision IP cameras, analog DVRs, and consumer-grade NVRs — that is typically used only for passive recording. The gap between “cameras that exist” and “cameras that contribute to operational intelligence” is substantial and rarely addressed in the face-recognition literature, which assumes greenfield IP deployments and large-scale labeled datasets. We describe a two-site pilot that retrofits face recognition and multi-camera worker re-identification onto unmodified legacy infrastructure at two distinct sites in Haryana, India: (i) a 22-camera Hikvision IP deployment at a gasket manufacturing facility (Phase 2, 3 cameras online, 3 workers enrolled), and (ii) a 16-channel analog DVR deployment at an Ayurvedic-pharmaceutical company office (12 cameras live, production face-recognition operational via CompreFace + Double Take + Frigate). The technical contribution is punch-seeded enrollment: a single timestamped biometric-attendance snapshot, already captured by the plant's turnstile, is used as the sole enrollment seed from which the system propagates worker identity labels across downstream camera footage via density-based clustering (DBSCAN) with a temporal window. This reduces annotation effort from “tens of labeled frames per worker” to “one automatically-captured punch snapshot per worker”. The paper reports only aggregate deployment metrics — no face images, no embeddings, no worker identifiers — consistent with India's Digital Personal Data Protection Act 2023 requirements for biometric data. We also document legacy-integration engineering gotchas that consumed a meaningful fraction of deployment effort: H. 265 hardware decode failures on GTX 970M (compute capability 5. 2), OpenVINO failing silently and reverting to CPU TFLite, digest-authenticated ISAPI calls on Hikvision cameras, and DVR SSH access that resets at every reboot. A companion Python script (computeₐggregateₘetrics. py) that produced the published numbers is released alongside this paper. The underlying pilot dataset (83 files, 787 MB comprising face images and multi-camera video from three workers) is not released in accordance with DPDP Act 2023 obligations.
Vibhav Aggarwal (Mon,) studied this question.