TTD-Bench: Raw Measurement Logs for Time-to-Detection Decomposition of CCTV Weapon-Detection Pipelines (Phase 1: UCF-Crime latency pilot; Phase 2: SCVD event-level TTD)

This dataset contains the complete raw evidence set for the paper introducing TTD-Bench, a replay-based hardware-in-the-loop benchmark that decomposes the end-to-end time-to-detection (TTD) of CCTV weapon-detection pipelines into five exactly-accounted components (acquisition, transfer, inference, post-processing, alarm policy), measured across deployment configurations (local GPU, emulated edge, remote inference behind LAN/WiFi/4G network emulation). It comprises 3 625 raw per-run JSONL logs from two frozen measurement campaigns on the Czech national grid MetaCentrum (225 runs of a per-frame latency decomposition on three UCF-Crime Shooting clips with COCO-pretrained detectors; 3 400 runs of event-level TTD measurement on the SCVD dataset with weapon-detection checkpoints), together with derived per-run CSVs, aggregated master tables, survival-analysis outputs (Kaplan– Meier, Cox proportional hazards, log-rank), a CONSORT-style run-accounting figure, the frozen experiment configurations (including the cluster-side variants that reproduce the logged config_hash), frame-level weapon-onset annotations for SCVD created for this work, and the pre-registered analysis plans and failure-diagnostic protocol. No video or image data are redistributed: the logs contain only timestamps and detection bounding boxes, and annotations reference the source videos by path. UCF-Crime (Sultani et al., CVPR 2018) and SCVD (Aremu et al., 2023) must be obtained from their original distributors under their own terms. The README inside the deposit documents the JSONL schema (nanosecond monotonic timestamps, event types), all CSV columns and units, the exact script pipeline that reproduces every published number, and known limitations (in particular 869 zero-frame runs caused by a diagnosed UDP port-collision infrastructure failure, which are accounted for in the CONSORT figure and excluded by the documented n_frames_processed > 0 filter).