IVS Briefs
IVS / Edge AI Research Brief
Signals (8)
- 1
Key point
Crews spent Thursday covering the city’s 55 Flock cameras with nylon bags after Tuesday’s Prop A reject (already numbered 11-04). Mayor Nick Long asked staff to cover immediately rather than wait for the Nov 17 canvass; the Oct council resolution still requires remove or cover within 72 hours of certification. Flock’s Courtney Terlecki said a remote-decommission request is in. About $120k/yr, ~1 year left on the contract; Long: “Worst-case scenario, we’ll satisfy the contract and move on.” The $120k stays in PD. Chief Cliff Woitena: losing the leads will change how detectives work some cases. Daily News: ~54% against; HPM rounded 56 / 46.
Why it matters
Hardware is already dark; the remaining job is physical take-down after Nov 17, not another vote. Bid local VMS / in-car Axon against those poles now, and write the pickup quote to the canvass + 72h clock, not election night.
- 2
Anderson Co SC unanimous Ord 2026-054: no county funds, property, or resources for Flock; 10 cams, $30–40k/yr
Key point
Council gave unanimous third reading to Ordinance 2026-054 after a public hearing Nov 3: no Anderson County funds, property, or resources for Flock Camera Systems. Independent Mail: 10 cameras still operating, most at boat ramps and parks, $30,000–$40,000/yr. Councilman Greg Elgin: Flock has begun taking cameras down, no all-down date. The Sheriff’s Office said in October it was already phasing out Flock and the vote would not significantly change operations. City of Anderson, other municipalities, and private entities keep their own cameras. First reading Oct 12, second Oct 21. Default Flock retain 7 days (some older setups 30).
Why it matters
A county-funds-and-property lock, not a statewide ALPR ban. Do not bid Flock onto Anderson County ROW; city and HOA poles next door are still live. Quote a take-down against the 10 county units Elgin says are already coming down.
- 3
Key point
Five members had introduced a resolution to terminate, not renew, remove equipment, and purge data. A motion to hold two weeks for the audit failed. The terminate vote then failed (six no; Brittany Jones absent). Jones arrived; a motion for another vote passed; the second vote tied 6–6. Mayor Wade Kapszukiewicz breaks the tie at the Nov 17 meeting. Cameras stay live while the auditor’s review continues (auditor asked a 30-day extension). ≥4 cameras vandalized in the last week. Nick Komives is ready to vote it down; Cerssandra McPherson wants the audit first. Do not treat this as a terminate.
Why it matters
The live fleet is one mayor vote from unwind or stay. Hold the take-down / keep-and-audit quote until Nov 17; vandalism already means some poles are dark before council acts.
Source
13abc / WTVG; The Blade - 4
Key point
AAEON’s 100 × 72 mm Pico-ITX is the first Pico-ITX Core Ultra Series 3 (Panther Lake) board. Four 25 W PBP SKUs: Ultra 5 325 / 5 338H / 7 355 / X7 358H (16C, Arc B390 12 Xe, 50 TOPS NPU, combined 180 TOPS). Up to 32 GB LPDDR5X-7467, M.2 2280 PCIe 5.0 x4 NVMe, M.2 2230 E-Key, FPC PCIe 3.0 x4, dual Intel I226 2.5GbE, HDMI 2.1 4K60, 2× USB 3.2 Gen2, RS-232/422/485, 4-bit GPIO, TPM 2.0. Typical 54.7 W / max 79.8 W (358H+32 GB). Ubuntu 26.04 (kernel 7.0.0-14) + Win11. Operating −20 to +70°C. Mass production; AAEON eShop in stock: $1,288 (325) / $1,456 (355) / $3,065 (358H); 338H unlisted. $33 CPU cooler recommended. Same 358H as LattePanda Sigma 2 (already 10-29); this is a new industrial Pico-ITX SKU with a price.
Why it matters
A priced, Linux, −20 to +70°C Pico-ITX you can drop behind a camera or in a small NVR, with model-swap headroom the locked ALPR camera does not have. Accelerated compute beats locked “AI cameras” (swap models vs a 36–84 month refresh).
- 5
Key point
Two months after Atlas (autoregressive world model, pixel-perfect camera control), World Labs opened a research preview: Frameboy streams frames from cloud GPUs to the browser at up to 15 fps as the user drives the camera. Integrated as a streaming overlay in Viewfinder (closed beta). Atlas generates from a single image or multiview; last year’s RTFM interpolated between existing viewpoints, Atlas generates from scratch. Cloud GPU, “not cheap compared to local rendering.” AMD deal already numbered 09-29.
Why it matters
Spatial replay for perimeter / construction / robot-nav demos is now an interactive 15 fps stream, still a cloud-GPU SKU. Quote it as a scene-preview tool, not an on-prem VMS overlay, until a local runtime ships.
Source
World Labs - 6
Key point
Zeping Zhang et al. argue 4D occupancy world models waste compute disentangling ego-motion the planner already knows. FreeOcc factors SE(2) self-motion out of a conditional flow-matching model so the ODE runs only on residual agent kinematics; an uncertainty-aware blender sends well-observed voxels to a single-pass decoder. 28.3M params (~40% of comparables). Ensemble FreeOcc-E at 27 FPS. nuScenes: 0.47 m Near Field Chamfer Distance, 32.1% mIoU; zero-shot SemanticKITTI; 92.8% of ideal mIoU under packet loss and timestamp jitter.
Why it matters
Occupancy RFPs fail on ego-motion and dropped packets before they fail on backbone size. Ask whether the model factors planner motion and still holds mIoU when the LiDAR stream jitters, not how many attention layers it has.
Source
arXiv:2611.08128 - 7
Key point
Kotaro Oishi et al. (Keio) run Transformers on full-waveform LiDAR histograms (fog, rain, glass, low-reflectance) instead of discrete returns. FWL-ToPM prunes/merges tokens from intensity so background rays drop; FWLAug keeps ToF waveform and frustum geometry. On Ghost-FWL: >8× speedup at real-time throughput and ~9-point F1 over prior SOTA — first real-time FWL Transformer that also gains accuracy. Code and an OOD unseen-scene set on the project page.
Why it matters
Perimeter LiDAR in rain/fog is still a histogram problem. Ask the integrator whether the stack keeps the full waveform and token-prunes low-intensity background, or discards it at the discrete-return driver.
Source
arXiv:2611.07613 - 8
Key point
Obed Honour Eje et al. benchmark YOLOv8n / YOLOv12n / RT-DETR / Faster R-CNN under physically grounded darkness, rain, and Harmattan dust. Uniform SCI low-light enhancement is harmful at dusk (−16.1% to −30.7% mAP@50). A training-free luminance-adaptive router (Rec. 601) picks none / moderate SCI / denoise+strong SCI per frame, recovering most of the dusk penalty while keeping night gains. Full pipeline validated on NVIDIA Jetson Nano. Paper, not a shipping SKU.
Why it matters
Always-on low-light enhance is a dusk regression. Quote a luminance router in front of YOLO/RT-DETR on Jetson-class boxes, and measure mAP at dusk, not only at night.
Source
arXiv:2611.08393