Blog

• Automatic Labels Need an Abstain Button September 28, 2026

  Why auto-labeling matters, how Lynceus combines tiled discovery, conservative naming, boundary refinement and verification, and what the measurements say it can—and cannot—claim. Updated with a train-on-the-labels test: auto-labels reach 91% of human.

• Measured, Not Reported: Four Edge Trackers on Anti-UAV410 September 25, 2026

  SAM 2.1, EfficientTAM and two UETracks on 129,691 thermal drone frames. Three of them lose the drone equally often; what separates them is whether they get it back — and when it leaves the frame for long, none does. With Jetson Orin NX latency for each.

• A Local Jira for Your Agents August 13, 2026

  Ten tasks, ten editor windows, sometimes three machines — and no agent can see the others. A small SQLite issue tracker they all share, with a board, a daily digest, and a scanner that checks their work.

• Quantization Is Three Problems, Not One July 25, 2026

  PTQ finds the ranges, QAT moves the weights into them, and QDQ translation hands them to TensorRT. Three stages that fail identically from the outside — and how to tell which one broke.

• When the Detector Disagrees With the Label June 24, 2026

  A deep dive on label integrity: flip TIDE error types from a model scorecard into a ranked label-audit queue. A detector disagreement is a suspicion, never proof.

• Active Learning Needs Orthogonal Axes June 24, 2026

  Not every model failure is a labeling problem. Why a single global active-learning score hides the decision — and how separating label repair, coverage, boundary uncertainty, domain shift, and model tuning into orthogonal axes makes data improvement precise and auditable.

• The Future of Edge Tracking: 2026 SOTA Single Object Tracking Review June 24, 2026

  Which 2026 tracker wins the accuracy–latency–stability trade-off under a fixed edge budget — accuracy-first RGB SOT, edge-first bbox SOT, and mask-based Track-Anything, compared on real Jetson runtime.

• What CVPR 2026 Says About Production Computer Vision June 12, 2026

  A practical review of what can ship now, what is near, and where research still outruns deployment.

• The Pipeline Is Moving Inside the Model June 13, 2026

  Software 2.0 replaced vision algorithms. The next transition is replacing the interfaces between them.

• The Map of CVPR 2026 June 7, 2026

  A data-driven tour of all 2,672 CVPR 2026 papers — an interactive map of the conference, the topics that dominate it, the ideas everyone is building on, and how the field names its work.

• Starting a notebook on edge perception June 7, 2026

  Why I’m writing here, and the kind of notes to expect — short, practical pieces on real-time detection, SLAM, and squeezing models onto edge hardware.