IVS Briefs
IVS / Edge AI Research Brief
Signals (8)
- 1
Key point
League City voters rejected Proposition A, the nonbinding question on whether police should keep using fixed ALPRs on public roads: 25,021 against (~54%) vs 21,412 for (~46%), unofficial. The city runs 55 Flock cameras at about $120,000/yr. Mayor Nick Long: “There’s no disagreement that they work. However, there’s also no disagreement that there’s serious privacy concerns.” Council’s Oct 13 resolution binds staff to remove or cover cameras within 72 hours of the official canvass; City Manager John Baumgartner is directed to review the Flock agreements. The Citizen Data Privacy Committee does one final audit, then dissolves. Long: the $120k stays in the police budget. First Texas city to put Flock on the ballot. Galveston County returns complete; Harris County incomplete. Clock starts at certification, not election night.
Why it matters
A ballot-forced 55-camera unwind with a 72-hour post-canvass SLA already on the books. Bid local VMS / in-car Axon against those poles, and write the take-down quote to the canvass date, not Nov 3.
- 2
Key point
York residents passed the citizen-petition “Stop ALPRs” ordinance 5,386–2,043. PD already terminated Flock in September and removed at least four cameras on public property after a 115-signature petition. Tuesday’s vote makes the ban permanent: no public agency in town may use similar technology. Press Herald notes it is already illegal for public entities under state law. The article is one of six local referenda; ambulance funding and a single-use plastics permanence question also passed.
Why it matters
New England’s ballot is now a town-code lock, not a PD non-renew. Do not bid ALPR into York. In nearby Seacoast towns still on Flock, expect the same petition-to-ordinance path the Maine Monitor flagged Monday.
Source
Portland Press Herald - 3
Key point
ZOTAC published full specs for three fanless ZBOX C-series nano boxes: CI471 (Core 7 350, 17 TOPS NPU), CI451 (Core 5 320, 16 TOPS), CI431 (Core 3 304, 15 TOPS). Shared 204 × 129 × 68 mm chassis, 15 W base / 35 W turbo, one DDR5 SO-DIMM up to 64 GB, M.2 Gen4 NVMe, HDMI 2.0b + DP 2.0 + dual Thunderbolt 4, 2.5GbE, Wi-Fi 7 / BT 5.4, VESA mount, 19 V / 65 W. Barebones; Windows 11 drivers posted; Linux not mentioned. Shown at Computex 2026; pricing and availability still unpublished. CNX notes AAEON UP WCL and BIOSTAR EdgeComp MS-U320 already ship the same Core 7 350 for industrial, while this SKU is consumer-oriented.
Why it matters
A VESA fanless NPU box you can hang behind a monitor or in a small closet VMS, with model-swap headroom the locked ALPR camera does not have. Quote it only after Linux and a price land; accelerated compute beats locked “AI cameras” (swap models vs a 36–84 month refresh).
Source
CNX Software - 4
Key point
Minghe Wang et al. (NeurIPS 2026) argue feed-forward VAU fails when predefined frame/event sampling misses the evidence, because the model cannot go back and look. SEEK-VAU treats anomaly understanding as a bounded search-and-verify loop over the three stages of an event chain — precursor, trigger, aftermath — and stops only after a sufficiency check. EFCV (evidence-faithful counterfactual verification) rewards evidence that stays supportive, compact, and necessary under counterfactuals. SEEK-Bench adds video-level episodes with interval labels, semantic QA, and stage tags.
Why it matters
Today’s VAD SKUs still fire on sampled clips. Ask vendors whether the model can re-search the precursor and aftermath under a token budget, or whether a missed frame is a missed case.
Source
arXiv:2611.04554 - 5
Key point
Qiao Li et al. (NeurIPS 2026) match UAV stills to ground CCTV. ProAlign adds Layer-wise Progressive Positional Embedding (LPPE) — conditional positional encoding in shallow layers, static embeddings deeper — and View-Aware Prototype Contrastive Learning (PCL) with per-identity aerial and ground prototypes. On CARGO plus two real AG-ReID sets: +1.60% Rank-1 and +5.01% mAP over the strongest prior.
Why it matters
Drone + fixed-camera ReID is showing up in perimeter RFPs. Quote prototype-per-view alignment when the UAV gallery and the gate camera disagree on pose and occlusion.
Source
arXiv:2611.04980 - 6
Key point
Aman Upganlawar trains a 3D detector only on synthetic clouds and tests on unseen real roadside LiDAR from an Ouster OS1-64 (UrbanTwin Sim2Real, V2X-Real track). Final 0.7457 combined (0.4382 mAP, 0.9251 realism), 1st place. Three diagnostic fixes, not a new net: coordinate-frame recovery (0.3520 → 0.5147), per-class NMS because a global 0.3 IoU on 0.7 m pedestrian boxes was killing grouped people (mAP 0.0751 → 0.4295), and modelling the ~31% real null returns encoded as (0,0,0) that synthetic clouds omitted.
Why it matters
Roadside LiDAR RFPs fail on frame, NMS, and null-return mismatches long before they fail on architecture. Ask the integrator how they handle OS1-64 zeros and per-class NMS, not which backbone they trained.
Source
arXiv:2611.03940 - 7
Key point
Yifan Liu et al. formalize Counterfactual Spatial Imagination (CSI) — predicting a scene after a hypothetical motion — and release FlowBench: 428k examples, sim + real, household and tabletop, with automatic optical-flow and outcome frames in simulation. Proprietary VLMs show an imagination gap (worse when the outcome must be inferred than when it is shown); open-source models sit near chance. Fine-tuning Qwen3-VL-8B on FlowBench-Sim lifts accuracy +29.5 pp in sim and +16.9 on real images. World-model text-to-flow: Psi-0.6 52.1% → 61.7%. Code and data released.
Why it matters
Occupancy and “what happens if that person steps into the lane” prompts are the next VLM add-on. Score the vendor on motion-outcome questions, not caption quality.
Source
arXiv:2611.03937 - 8
Key point
Siqiao Huang et al. argue video-generation tokenizers keep perceptual clutter that world models do not need. Compression Forcing adds Dynamics-Aware Tokenization (DAT) — history ordered by decreasing dynamics saliency — and Hierarchical History Guidance on a diffusion backbone so inference can amplify the moving bits. Gains reported on prediction fidelity, physical realism, and downstream control.
Why it matters
Long-video agents on a Jetson or NPU box cannot keep every background pixel. Prefer a tokenizer that ranks what moved when you spec a streaming world-model sidecar.
Source
arXiv:2611.04028