IVS News

IVS Briefs

IVS / Edge AI Research Brief

Signals (9)

  1. 1

    Taylor, TX: PD terminates Flock; 8 cameras on Hwy 79 / 95 already disabled

    Key point

    Taylor terminated its Flock Safety contract for eight ALPRs on Highway 79 and Highway 95, in place since late 2023. Police Chief Joseph Chacon: frequency of use was “far less than anticipated” and usefulness “has not been realized in solving crimes,” and the city is not satisfied with Flock data-security and protection measures. “We may revisit the use of Automated License Plate Readers in the future should data security concerns be addressed.” The 8 cameras are already disabled and “will be removed in the coming days.” The city does not maintain any other ALPRs.

    Why it matters

    A PD-led Texas terminate on usefulness plus security, not a council privacy fight. Bid a local VMS against eight highway poles that are already dark, with an open door if a vendor can show data custody the chief will sign.

  2. 2

    Two Rivers unanimous Flock terminate: 12 cameras bagged, Flock pickup spring

    Key point

    Two Rivers City Council voted unanimously Monday to terminate Flock. Cameras were already deactivated in September and data-sharing was cut then; the new motion keeps them covered until removal. Police Chief Ben Meinnert (also acting city manager): “It’s just wasted money when we are not using it as it is originally intended.” He still wants strategically placed cameras but will wait for federal guidance before another contract. City attorney: a clean break, but Flock will not pick up until spring. Herald Times Reporter (Sept. 15): the city runs 12 Flock cameras. Northeast Wisconsin unwind already includes Appleton (vendor-class ALPR limits), Oshkosh, Grand Chute, and Kaukauna.

    Why it matters

    Lakeshore Wisconsin is now a bag-until-spring unwind, not a same-week hardware-down. Quote a 12-camera cover plus a federal-guidance hold before any Axon/Genetec swap.

  3. 3

    Renton lets 24-camera Flock contract expire Nov 2; Axon in-car and Genetec parking stay

    Key point

    Renton’s two-year Flock deal ended Nov 2. Mayor Armondo Pavone confirmed non-renewal at the Oct 26 council meeting; Councilmember Ruth Pérez said city attorney Shane Moloney told council Oct 21 that police had sent Flock a notice of non-renewal so auto-renew would not fire. 24 stationary Falcons at major intersections, contracted Nov 2024, about $80,000/yr (~$170,000 over two years). Cameras have been off since May 4 and covered since Aug 3. Pavone: they come down after expiration. Axon in-car ALPRs and Genetec parking-enforcement readers were turned back on July 22. Private Flock on private property can stay; PD does not map those sites.

    Why it matters

    A lapse-and-remove with the in-car Axon / Genetec parking stack already back on. Bid the 24 intersection poles as a local VMS, and do not treat the city as ALPR-free.

  4. 4

    Henrico: 100+ Public Works crews self-removed 120+ Flock cameras after Flock missed Oct 21

    Key point

    Henrico notified Flock Oct 7 that the contract ended Oct 14 and cameras had to stop collecting that day, with a remove-by Oct 21 ask. When Flock missed it, more than 100 Public Works employees spent Saturday working from both ends of the county and took down more than 120 cameras (Henrico Citizen Oct 29: 128 still standing, already deactivated). Cameras go to county storage until Flock retrieves them. Poles stay; Flock owes pole removal and the county’s take-down cost. County Manager John Vithoulkas told WWBT Flock did not meet its “contractual obligation.” Flock: cameras stopped capturing on terminate; standard terms are “commercially reasonable” time; the county self-removed; Flock says it has reached out again. County Attorney Andrew Newby Oct 29: no replacement ALPR buy in motion.

    Why it matters

    Largest self-remove of the cycle: 120+ heads in storage, poles still in the ROW, vendor disputing the clock. Write the RFP as city-held pickup + pole restoration + cost-shift, not a Flock SLA.

  5. 5

    Maine Monitor: Flock enlisted York DS Thomas Cryan and other officers to sell ALPRs to neighboring towns

    Key point

    Emails obtained by the Monitor and ACLU of Maine show Flock salesman Dan DelGrosso used York Detective Sergeant Thomas Cryan as a regional closer: Cryan sat in on pitches to Eliot and Sanford chiefs, talked up Flock to the Cumberland County Sheriff, flagged Kennebunk’s chief’s vacation, and introduced a motel owner plus a parks-board ex-chief ($11,000 Short Sands Beach proposal, not bought). Freeport Detective Beck Rice got “prep calls” with Flock’s municipal-engagement director; in Feb 2026 DelGrosso offered Freeport Project Prove It — one or two months free, no commitment. Wells then-Chief Jo-Ann Putnam and Town Manager Michael Pardue pitched the select board with DelGrosso after a vendor “presentation review.” TCU’s Johnny Nhan: users say Flock is “like Costco or drug dealers” — free sample, then contract. Flock did not comment. York’s Stop ALPRs citizen ordinance is on the Nov 3 ballot; cameras there are already covered.

    Why it matters

    Officer-as-channel and free-trial footholds are the bid you are actually competing with. Ask the city attorney whether a no-trial / no-officer-pitch clause is in the procurement, especially in New England towns still deciding after York.

  6. 6

    DADKC: cloth-hybrid lifelong ReID without storing old images; +4.1 / +4.5 mAP/R1

    Key point

    Yan Zhang et al. treat Cloth-Hybrid Lifelong Person Re-identification (CH-LReID): cameras and outfits both shift over time, and you cannot keep old images. DADKC adds three modules on Differentiated Knowledge Consolidation — Anchor Pool Adapter, Phase Path Mixer, Stage Bridge Alignment — and stores neither historical frames nor per-image features. On CH-LReID: +4.1% / +4.5% mAP/R1 on Order-1 and +3.6% / +3.4% on Order-2 vs DKC; weighted forgetting drops from 2.4% / 1.7% to 1.8% / 1.4%. Accepted MMM 2027. Code: github.com/fire667/DADKC.

    Why it matters

    Multi-site ReID that has to survive seasonal clothes and new cameras without a replay store. Quote DADKC-style consolidation when a VMS ReID add-on cannot legally keep last year’s gallery.

  7. 7

    VLM-as-a-Judge: frozen VLM rerank beats dedicated ReID experts +12.9 rank-1 on cloth-change

    Key point

    Bessie Dominguez-Dager, Miguel Cazorla, and Vincent Lepetit use a zero-shot general-purpose VLM as a judge that reranks a short ReID candidate list. On all four CHIRLA long-term cloth-changing scenarios (no public ReID model trained on it), reranking two frozen experts lifts rank-1 by 12.9 points over the best expert and 9.4 over a naive ensemble, with nothing fine-tuned on the target. Invoking the VLM only where the two experts disagree at rank 1 drops 40% of VLM calls (and a similar share of FLOPs) with no significant rank-1 drop, trading 1.8–3.8 mAP. Almost all pipeline FLOPs still sit in the judge. NeurIPS 2026 VLM4RWD workshop. Code: github.com/bdager/JudgeReID.

    Why it matters

    A shortlist + VLM judge is a cheaper path than retraining ReID every time outfits change — gate the VLM on expert disagreement so the GPU is not on every track.

  8. 8

    StreamClarify: streaming VLM asks when the query is ungrounded; +8.2 Macro accuracy

    Key point

    Zhuchen Shao, Bo Chen, and Klara Nahrstedt (UIUC) attack streaming video QA: current streaming VLMs answer confidently when the query is ambiguous, has a false premise, or the evidence has already left the active window. StreamClarify is training-free: entropy of a binary pre-answer check decides answer vs clarify; a temporal-pyramid KV cache holds bounded live context and a CPU archive retrieves older frames. StreamClarify-QA: 1,837 pairs (standard / referential ambiguity / unanswerable). Across three VLM backbones, Macro accuracy rises by up to 8.2 points; entropy gating is >18 points more precise than random at catching initially wrong answers. WACV 2027.

    Why it matters

    Operator-facing video agents that ask instead of hallucinate when the person left the frame. Spec a clarify-or-abstain gate on any streaming VLM RFP, not a forced answer.

  9. 9

    EvoOcc: fast-slow occupancy so the cheap model holds the scene until the accurate one lands

    Key point

    Runhe Yang et al. (NeurIPS 2026) fuse a fast occupancy net with a slower, more accurate one as a continuous-time controlled dynamical system (EvoOcc). Ego-state evolution undoes known coordinate shift; scene-state evolution models real geometry/semantics. It does not replace the two upstream models. On three fast-slow pairs and irregular arrivals, EvoOcc beats the fast baseline while staying much lower-latency than the slow system.

    Why it matters

    Perimeter occupancy RFPs already pair a cheap edge pass with a heavier fusion box. Quote EvoOcc-style fast-slow evolution so the live occupancy grid does not freeze while the slow model is still in flight.

Continue

Directory

LPR/OCR