Field note · 01

Where generic recognition fails

Where current object detectors excel, where they fail, why those failures cluster, and what we design against. Public sources; vendor figures treated as directional.

Most recognition failures are predictable. They cluster around a handful of causes: the scale of the object, the shift between the world a model was trained in and the one it is deployed in, clutter, concealment, the physics of the sensor, and the link back to the operator. This note sets out what the public record says about each, and what we design against as a result. It was compiled on 8 October 2026 from public, secondary and vendor sources. Treat vendor figures as directional.

Where the field stands

Two architecture families dominate. Grid-based single-stage convolutional detectors, the YOLO line, remain the default for speed on small hardware. Transformer detectors (DETR and its real-time descendants, RT-DETR and RF-DETR) draw on global image context, and their proponents report fewer false positives in cluttered scenes: one independent outdoor vehicle deployment reported a false-positive reduction of roughly half against a YOLO baseline, a figure we treat as directional rather than general.

Open-vocabulary and foundation models reduce the labelled data needed to start and speed up fine-tuning, but a 2026 ICLR paper reports that they still degrade under distribution shift and on categories they were not shown.

Edge inference is routine. Jetson Orin-class modules now run detection and optical navigation on small drones fielded by both sides of the war in Ukraine.

Licensing shapes choice. The current YOLO release is under the AGPL with a paid enterprise licence; small RF-DETR variants are under Apache 2.0.

Nine ways detectors fail

Small objects

Between 60 and 80 per cent of the objects in drone imagery occupy fewer than 32 by 32 pixels, with low contrast and sparse features. General multimodal models miss many small vehicles in satellite tests. We build multi-scale model layers and track across frames, so a vehicle that is faint in one frame is confirmed over several, and we cue higher-resolution collection to confirm what wide-area search finds.

Domain shift

A satellite benchmark reports a 35 to 40 per cent accuracy drop when a detector moves to an out-of-distribution climate zone. A 2026 study of transformer detectors finds that as the shift grows, missed detections dominate the error, not false positives. Convolutional detectors overfit to local cues such as lighting and texture; the positional queries of transformer detectors are sensitive to image style. We train and evaluate on the customer’s conditions, evaluate every release per class and per condition, and treat missed detections as the primary risk metric.

Clutter and false positives

Grid-local reasoning reads shadows, rubble and scrap as vehicles, and the choice of threshold drives false positives and false negatives as much as the architecture does. We use global-context architectures, calibrated confidence and an explicit unknown class, so the system says what it does not know.

Camouflage, concealment and decoys

Nets, treelines and dazzle paint are designed to defeat both eyes and models, and a 2026 adversarial-camouflage study targets optical and infrared detectors together. One vendor reports vision models that track through shadow and treeline (a vendor claim). An analysis from CSIS finds that smaller specialised models outperform one large model on drone footage and help with false-positive assessment. We fuse electro-optical and infrared inputs where they exist, reason over shape and context rather than texture alone, report a decoy likelihood as an output, and keep model layers specialised per domain.

Thermal contrast

Infrared defeats foliage and light fog but fails when the target matches its background, and active thermal camouflage can mimic benign signatures; multispectral fusion is the recommended answer for detecting the anomaly. We learn per-band signatures and require cross-band agreement, so a loss of contrast in one band does not become a loss of the target.

Weather and motion

Fog, motion blur, low frame rates and compression degrade aerial video. We augment for blur and compression and rely on temporal tracking rather than single frames.

Jamming blankets navigation and video links, and the main battlefield value of onboard vision is that it keeps working after the link is lost. We run recognition fully on edge compute and hold results locally, with the human decision preserved and logged.

Data scarcity

Military imagery is confidential, and researchers resort to simulation and augmentation. We work with customer data under controlled handling and use synthetic augmentation where it is warranted.

Fine classification

A Ukrainian developer describes current systems as able to separate a tank from a person “but not more than that.” We build hierarchical taxonomies down to type, variant, configuration and markings, defined with the customer.

What the battlefield demands

Ukraine’s defence ministry says more than seventy AI and computer-vision systems are striking targets and that it wants computer vision on every front-line drone, with a human retaining the final engagement decision. “Middle strike” drones fly under operator control to a target grid, then onboard vision takes over navigation and targeting through jamming. One vendor reports a brigade’s hit rate rising from a fifth to four fifths with last-mile vision lock-on; the figure is a vendor claim and unverified. An acoustic detection network feeding a common operating picture reports a false-positive rate near 1.6 per cent, which is a reconnaissance figure, not a measure of terminal-seeker decoy rejection. We found no public, independent test data on terminal-seeker false positives or decoy rejection.

What this means for how we build

  1. The honest differentiator is not “more accurate” but engineered for the failure modes: scale, shift, clutter, concealment, sensor physics and denied links.
  2. Hierarchical identification, from platform to type, variant and markings, is a real gap in fielded systems.
  3. Preserving the human decision and keeping evidence records is both a selling point and a compliance posture.
  4. We make no quantitative accuracy, range, latency or zero-miss claim on this site. A future claim will name the evaluation, the sample, the configuration and the operating scope, with its uncertainty.

Sources

  1. Roboflow, “RF-DETR vs YOLO”: https://blog.roboflow.com/rf-detr-vs-yolo/
  2. RT-DETR paper: https://arxiv.org/html/2304.08069v3
  3. JetBrains PyCharm blog, best object detection models for machine learning in 2026: https://blog.jetbrains.com/pycharm/2026/07/best-object-detection-models-for-machine-learning-in-2026/
  4. Benchmarking RF-DETR against YOLO in a real-world system: https://medium.com/@pratikmachchar/benchmarking-rf-detr-against-yolo-in-a-real-world-system-351a16d229a9
  5. Satellite domain-shift benchmark: https://arxiv.org/html/2503.19202
  6. Transformer detectors under distribution shift (2026): https://arxiv.org/html/2608.24154v1
  7. ICLR 2026, open-vocabulary detection under shift: https://proceedings.iclr.cc/paper_files/paper/2026/file/6d998a2ae756f71fe3b6fb29bf015a45-Paper-Conference.pdf
  8. Small objects in drone imagery: https://arxiv.org/pdf/2211.07483
  9. Multispectral detection of thermal camouflage: https://www.nature.com/articles/s41598-025-99634-0
  10. Adverse weather and UAV video degradation: https://www.mdpi.com/2076-3417/12/23/12236
  11. Adversarial camouflage against optical and infrared detectors: https://arxiv.org/html/2507.05229v1
  12. Aerial detection survey: https://link.springer.com/article/10.1007/s44443-026-00542-8
  13. Euromaidan Press, AI in Ukraine’s long-range strike system: https://euromaidanpress.com/2026/06/12/how-ukraine-is-integrating-ai-into-its-long-range-drone-strike-system/
  14. Kyiv Independent, autonomising drones: https://kyivindependent.com/ukraine-is-autonomizing-more-of-its-drones-ai-is-only-part-of-the-solution/
  15. Defence Industry Europe, seventy AI-enabled systems: https://defence-industry.eu/ukraine-says-more-than-70-ai-enabled-systems-are-striking-battlefield-targets-as-it-seeks-computer-vision-across-all-frontline-drones/
  16. CSIS, Ukraine’s future vision and current capabilities: https://www.csis.org/analysis/ukraines-future-vision-and-current-capabilities-waging-ai-enabled-autonomous-warfare
  17. IEEE Spectrum, autonomous drone warfare: https://spectrum.ieee.org/amp/autonomous-drone-warfare-2676377272
  18. Asia Times, the race toward fully autonomous drones: https://asiatimes.com/2026/09/ukraine-war-speeding-lethal-race-toward-fully-autonomous-drones/
  19. Tom’s Hardware, Jetson Orin in a next-generation strike drone: https://www.tomshardware.com/tech-industry/artificial-intelligence/russia-allegedly-field-testing-deadly-next-gen-ai-drone-powered-by-nvidia-jetson-orin-ukrainian-military-official-says-shahed-ms001-is-a-digital-predator-that-identifies-targets-on-its-own
Built around your challenge

What do you need
to recognize?

Start a project