IVS Briefs
IVS / Edge AI Research Brief
Signals (9)
- 1
Key point
New Bedford City Council passed Councilor Shane Burgo’s ordinance 8–3 Thursday barring city departments from acquiring, purchasing, leasing, licensing, renewing, expanding, deploying, operating, or materially modifying ALPRs and mass location-tracking. Nos: at-large Naomi Carney, Brian Gomes, Ward 3 Shawn Oliver. WBSM: the 8–3 is veto-proof if Mayor Jon Mitchell rejects it. The same day Mitchell fired Officer Emily Pacheco and Sgt. James Marinelli for personal Flock searches (Light: Pacheco’s former partner more than 650 plate hits). PD confirmed to The Light Friday that Deputy Derek Belong sent Flock written non-renewal on Oct. 8, more than 30 days before the Nov. 12 expiration. Cameras have been off since Aug. 11. Mitchell is developing an alternate ordinance; Ordinance Committee is today Oct. 26. Burgo: “The door is not shut on ALPR technology.”
Why it matters
The hardware path is a Nov. 12 non-renew with cameras already dark. Do not bid Flock back in; a guarded local-retain ALPR stack is only live if Mitchell vetoes and council later reopens Burgo’s ban.
- 2
Southgate, KY 3–2: cover 5 Flock cameras until the 2027 General Assembly; contract paid through March 2027
Key point
Southgate City Council voted 3–2 Wednesday to suspend and cover its five Flock readers while Kentucky lawmakers consider ALPR rules in the 2027 session. Yes: Liz Hartman, Mark Messmer, Aileen Okura. No: Paul Melville, Tom Whitehead. Jackson Curry absent. Five cameras, about $15,000/year, contract already paid through March 2027. Council covered instead of terminating to avoid a $500/camera ($2,500) pickup fee. If the 2027 session produces no statute, the item returns to council. Chief John Christmann defended the network; Southgate was the first Northern Kentucky PD on Flock.
Why it matters
This is a paid-through cover, not a take-down. Do not quote the readers gone; a March 2027 pickup or a 2027-session reopen is the live calendar.
Source
LINK nky - 3
Key point
A motion to renew Lake Jackson’s $12,500/year Flock payment failed 3–3 Monday, ending nearly four years of city ALPRs. For renewal: Ralph Buell, Jon “J.B.” Baker, Kevin Herrington. Against: Mayor Rhonda Seth, Leonard Schooler, Matthew Bjune. City Manager Modesto Mundo: the city will not renew and the matter will not return to council. The city must notify Flock by Nov. 6; Chief Kibodeaux said Flock will pull the five cameras when the contract ends in December. Kibodeaux recommended renewal and said the department had no local abuse complaints.
Why it matters
This is a failed-renewal take-down with a Nov. 6 notice clock. Do not bid city Flock in Lake Jackson; December pickup is the residual work.
Source
The Facts - 4
Key point
Rockchip’s upcoming RK3668 sits above RK3588 (RK3688 remains the planned flagship): 6 nm, six Cortex-A725 to 2.5 GHz plus two Cortex-A520 to 2.2 GHz, Armv9.2 / SVE2, Mali-G2 Pro MC4 (~1.23 TFLOPS FP32 at 1.1 GHz). NPU: 44 TOPS sparse INT8, INT2–INT16 plus FP4/6/8/16, CNN / Transformer / MoE. Memory: LPDDR5/5X-8533 (64-bit) or LPDDR6-10667 (96-bit), up to 64 GB, ~90 GB/s effective (128 GB/s peak). Video: decode 8K120 (H.265/H.264/VP9/AV1/AVS2), encode 8K60 H.264; two 8K60 or six 4K60 displays. Expansion: PCIe 5.0, 10 GbE, SATA 3.1. Rockchip lists NVRs, robotics, and industrial boxes among targets. Not a shipping module this week.
Why it matters
Quote RK3668 when an NVR or multi-cam box needs 8K decode plus a 44 TOPS NPU without a locked camera SoC. Accelerated compute beats locked AI cameras (swap models vs a 36–84 month refresh).
Source
LinuxGizmos - 5
Key point
Mulugeta Weldezgina Asres, Lei Jiao, and Christian Walter Omlin present XAVAD: zero-shot anonymization plus anomaly-type categorization, Grad-ECLIP spatial heatmaps (prompt-guided patch autofocus), and VLM scene/reasoning text on top of CLIP video embeddings with LLM-assisted prompts. Evaluated on public CCTV sets; they report substantial gains versus prior XAI-for-VAD baselines and publish privacy-protection trade-offs. Code: github.com/muleina/XAVAD.
Why it matters
This is VAD that can talk and hide faces, which is the RFP language privacy offices ask for when a city will not accept a raw identity stream.
Source
arXiv:2610.32073 - 6
Key point
Khaled Waleed Dawoud and Muhammad Haris Khan introduce SAAF: a dual-signal sampler in embedding space (frame-to-frame delta plus rolling-context drift, OR-gated with adaptive EMA thresholds) feeding a threshold-gated cascade that only spends video-level VLM compute when frame scores warrant it. Training-free AUC-ROC: 92.07% XD-Violence, 86.70% MSAD, 80.54% ShanghaiTech — SOTA among zero-shot methods; XD-Violence and MSAD also beat supervised baselines. Throughput 44.95 / 48.61 / 41.06 FPS on those three (2.22–2.63× vs prior best zero-shot). Code: github.com/kdawoud91/SAAF.
Why it matters
This is the VAD path that runs on a live stream without a site training set. Quote the FPS numbers when an MSP is asked to put anomaly on an existing NVR GPU rather than a new camera SKU.
Source
arXiv:2610.32486 - 7
Key point
Xiaoyang Li, Mingming Lu, Wei Zhou, Hao Li, and Canhui Zhang argue that dense text-residual fusion for visible–infrared ReID entangles useful appearance with redundant semantics. SARC (Superposition-Aware Residual Calibration) rebuilds a sparse residual over a shared dictionary, then calibrates it against identity-specific RGB–IR visual gaps. Improves the VI-ReID backbone on Tri-SYSU-MM01, Tri-RegDB, and Tri-LLCM versus treating textual differences as visual complements.
Why it matters
Dual-spectrum gallery search fails first when the text prompt and the IR crop disagree. SARC is the Lab path that filters the prompt residual before it hits the matcher.
Source
arXiv:2610.32709 - 8
Key point
Haohua Que, Mingkai Liu, Tianle Zhu, and Handong Yao move an ACEsplat-style scene-coordinate + 3D Gaussian pipeline onto the AMD KR260. Board-measured: INT16 SCR localization 25.42 ms at DSAC* accuracy; renderer 34.11 dB on Cambridge vs ACEsplat 33.20; RealEstate10K novel-view 27.84 dB vs 26.11. Wins performance-per-watt over a Jetson Orin NX on every deployment task (1.9×–5.7× at about a quarter of the power) and is faster in absolute latency on localization and Wayspots rendering. On-device training reaches ACEsplat quality but not GPU time or energy.
Why it matters
Live site models are moving off the desktop GPU. LENS is the Lab path that puts localization + novel-view on a low-power FPGA instead of a locked camera SoC.
Source
arXiv:2610.25068 - 9
Key point
Eugene Denteh, Blessing Agyei Kyem, Andrews Danyo, Joshua Kofi Asamoah, Richard Dzinyela, Solomon Owusu-Ansah, Dickson G. Asiedu, and Armstrong Aboah fuse six Qwen3-VL checkpoints with frozen SigLIP2 four-view gallery encoding, then rerank with Qwen3-VL-Reranker-8B. Training is synthetic-only; evaluation is real surveillance. Official test: 94.7815% mAP, 90.8493% R@1, eighth on the public leaderboard.
Why it matters
“Find the person in the red jacket who climbed the fence” is the query MSPs get after an incident. This is the Lab path that does that from text without a labeled gallery.
Source
arXiv:2610.31930