IVS Briefs
IVS / Edge AI Research Brief
Signals (9)
- 1
Key point
Troy City Council voted Thursday evening to override Mayor Carmella Mantello’s veto of Local Law 3. Spectrum did not publish a roll-call tally (the original Oct. 1 pass was 6–0 on an all-Democratic council). The law erases ALPR data after 72 hours unless the defense requests preservation. Mantello: the law is biased toward criminal defendants, undermines TPD’s ability to solve violent crime, and the council refused a working group with her administration, police, and the DA. Council President Sue Steele: veto was expected; the law is “an excellent compromise”; the mayor “was never willing to listen to the concerns of citizens.” Noreen McKee: “the mayor never read the law.” Katie Spain-McLaren: “We absolutely took the advice of the police.” Prior coverage of the statute: ALPR use limited to felony-linked vehicles, missing persons, stolen vehicles, uninsured/unregistered, city parking, and secured-area access; sharing only with a court order or the Rensselaer County DA. Cameras stay; this is guardrails, not a take-down.
Why it matters
Troy is a retain/search/share rider that survived veto, not a Flock exit. Bid against 72-hour delete, felony-tied search, and court-or-DA share; do not quote the readers dark.
Source
Spectrum News - 2
Key point
Delavan City Council voted 7–1 Tuesday night to cancel its Flock contract. City crews permanently took down all nine Raven license plate readers on Wednesday. The city will not install any similar products from competing companies in 2027. Nine Raven readers at $27,000/year. Petition: Rachel DeBow. Resident Jeff Terhune on the dais: “Hit the road, Flock.” TMJ4 does not name the no vote.
Why it matters
This is a same-week take-down plus a vendor-class 2027 lock. Do not bid city ALPR in Delavan through 2027; the hardware is already off the poles.
Source
TMJ4 - 3
Key point
Poughkeepsie Common Council passed R-26-66 Tuesday to end all Flock contracts (LPRs and gunshot detection) and require guardrails before any future ALPR. Mid Hudson: only Councilman Nathan Shook opposed. Sponsors Evan Menist and Daniel Atonna. Mayor Yvonne Flowers announced Wednesday she will veto, calling for a comprehensive review with police before stripping tools that also feed parking/traffic fines (Mid Hudson: more than $1 million a year). Contract runs through January 2027; a planned 2027 speed/red-light expansion is also in the resolution’s path. Dutchess DA and sheriff already dumped Flock last month. Cameras are not off.
Why it matters
This is a veto calendar, not a take-down. Watch the override math and the January 2027 contract end; do not quote Poughkeepsie readers dark.
- 4
Key point
County Executive Jen Metzger issued Executive Order 5 of 2026 prohibiting all departments and employees under her authority from purchasing, contracting for, or deploying Flock cameras or similar AI-assisted surveillance, including vehicle or person reidentification, and bars vendors/contractors from using those tools on the county’s behalf. Effective immediately until superseded or revoked. The county currently does not use these technologies. The order does not apply to the Sheriff. Sheriff Juan Figueroa supports the EO; the office has used Motorola ALPRs for more than a decade and says it will not use Flock. Kingston’s Jenoptik school-zone ALPRs are city, not county.
Why it matters
This is a preemptive exec ban, not cameras down. Do not bid Flock or ReID under county departments; Sheriff Motorola ALPR and Kingston school-zone Jenoptik remain the live channels.
- 5
Foundation4D: NVIDIA unified feed-forward geometry, 3D detection, semantic occupancy, Gaussian recon
Key point
NVIDIA authors introduce Foundation4D, a unified feed-forward model for geometry estimation, 3D detection, semantic occupancy, and Gaussian reconstruction from multi-camera driving video. A shared backbone uses factorized spatiotemporal attention; pose-graph optimization refines trajectories and assembles object tracks across windows. They release PAI-Foundation4D, dense depth and semantic occupancy labels on 1,253 driving clips. With benchmark-specific fine-tuning: lowest depth AbsRel on Waymo and nuScenes, highest detection mAP on nuScenes, highest occupancy mIoU on OpenOccupancy among compared methods. Co-training detection with geometry and occupancy lifts top-class mAP +1.2; co-training geometry with detection cuts camera-to-rig translation error 13.4%. Predicted geometry and trajectories feed lidarless NuRec for simulation.
Why it matters
Multi-cam occupancy plus 3D detection in one feed-forward pass is the Lab path into a camera-only perimeter / site-occupancy stack that does not wait on a LiDAR SKU.
Source
arXiv:2610.30340 - 6
Key point
Khaled Waleed Dawoud, Muhammad Zaigham Zaheer, Khan Muhammad, and Muhammad Haris Khan propose MoG-VLM: Temporal Gradient Magnitude (TGM) content-aware sampling plus a hierarchical dual-stream that runs dense multi-frame VLM reasoning only around elevated-anomaly frames, with early termination. Zero-shot, training-free AUC-ROC: 91.91% XD-Violence, 84.30% UCF-Crime, 87.88% MSAD, 78.96% ShanghaiTech — SOTA on three; +0.57% vs prior on XD-Violence at 1.31× inference speedup. Code: github.com/kdawoud91/MoG-VLM.
Why it matters
This is the VAD path that skips static frames instead of densifying VLM inference. Quote it when an MSP is asked for scene-description anomaly on a live camera without a site-specific training set.
Source
arXiv:2610.31478 - 7
Stream4D: training-free streaming 4D recon; causal KV-cache motion masks kill long-sequence ghosting
Key point
Wei Zhang, Ying Feng, Mengtong Wu, Shiqiang Gong, Huaqi Zhang, and Qi Zhang introduce Stream4D, a training-free motion-decoupled streaming reconstruction layer that drops into causal Transformer backbones (StreamVGGT, Stream3R). Causal attention’s KV cache is aggregated across layers and temporal windows, fused with depth priors, to produce dynamic masks that suppress moving regions in pose estimation. Soft key masking improves camera pose and static-scene quality and cuts long-sequence ghosting on dynamic-scene benchmarks.
Why it matters
Live multi-cam site models fail first on people walking through the reconstruction. A training-free KV-cache mask is the Lab path into streaming occupancy without a per-site fine-tune.
Source
arXiv:2610.31320 - 8
Key point
Ali Safaei and Mohammad Kazemi present MS2CAST, an end-to-end semi-supervised predictive VAD framework. Spatial encoding fuses MViTv2-S and ConvNeXt-T via multi-scale cross-attention; interchangeable latent spatiotemporal modules (SwinLSTM or VMRNN/Vision Mamba) replace optical-flow networks. Loss: L1 + GDL + MS-SSIM. Frame-level ROC-AUC on UCSD Ped2, CUHK Avenue, and ShanghaiTech is competitive with top-tier methods at peak GPU memory under 1 GB.
Why it matters
Sub-1 GB predictive VAD is the Lab path onto a small NVR or Jetson-class box without an optical-flow tax.
Source
arXiv:2610.29649 - 9
Key point
Oron Nir, Eliyahu Strugo, and Ariel Shamir propose MINER, a model-agnostic clothes-changing ReID framework that suppresses garment bias by contrasting hard examples and adaptively balancing dataset-dependent intra- and inter-person variance. SOTA on four CC-ReID datasets; among the leaders on all three tracks of the open TVRID RGB-Depth ReID challenge.
Why it matters
Gallery search after an outfit change is the ReID failure mode MSPs get asked about at the second camera. MINER is the Lab path that does not depend on a segmentation-mask heuristic.
Source
arXiv:2610.28941