Object Detection Patents: Who Leads, Where the Gaps Are 2026
- Concentration is modest. the top 5 assignees hold just 11.1% of all 733 records in scope, and the top 10 combined reach only 18.6% — this field has no single gatekeeper.
- Filing growth has flattened. annual filings ran from 24 in 2017 to a peak of 112 in 2025, with the 2022 midpoint at 103 — the recent trend is flat to declining, not accelerating.
- Recognition and AI-model classes dominate co-occurrence. G06V image/video recognition appears on 73.0% of records and G06N AI-model computing on 52.4%, meaning most filings claim recognition logic paired with a learning model, not one alone.
What the 733-record corpus covers
This landscape draws on 733 published records filed between 2015 and mid-2026 that combine object-detection search terms with claim-level language on mean average precision, small-object detection, inference latency, anchor-free design or label noise, restricted to IPC classes covering image recognition, AI-model computing and image data processing. That combination narrows the corpus to filings that go beyond generic computer-vision claims and specify a detection architecture or its evaluation criteria. Publication lags filing by roughly 18 months, so the 2026 count in particular understates real filing activity for that year.
Filings arrive through six receiving offices, with the United States and India together accounting for the largest share of records, followed by China, the EPO, WIPO's PCT route and Japan. The most-cited records in the set date from 2018-2019, consistent with citation counts favouring older filings that have had more time to accumulate references rather than signalling which architectures are most active today.
Filing trend and technology composition
Two views of the same 733 records: the pace of filing over time, and which IPC subclasses those filings sit in. Because a single record can carry several IPC classes, the composition shares add up to more than 100% of the record total — that is expected and does not indicate double counting.
Annual filings, 2017-2026
Filings rose from 24 in 2017 to a peak of 112 in 2025. The 2022 midpoint of 103 sits close to that peak, meaning growth flattened well before the most recent years rather than continuing to climb — and the 2026 figure of 61 is a partial year still subject to publication lag.
IPC subclass composition
G06V (image/video recognition) appears on 73.0% of the 733 records and G06N (AI-model computing) on 52.4%, confirming that most filings pair a recognition pipeline with a learning model rather than claiming either in isolation. G06T (image data processing) reaches 42.6% and G06K (data recognition/presentation) 27.4%, while adjacent domains — A61B diagnosis/surgery, H04N video communication and G01S radar/positioning — each sit near 3% of records, marking them as smaller but present application vectors rather than core territory.
Shares are the percentage of the 733 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Go deeper on Object Detection Architectures with Eureka
This page is one run against one query. Ask Eureka your own question about object detection architectures and every answer comes back with the patent numbers behind it.
Try EurekaRepresentative filing and most-cited records
Small object detection method and apparatus, readable storage medium, and electronic device
The disclosure describes a small-object detection method that separately encodes and decodes image information using paired desubpixel and subpixel convolution operations, then extracts features to output an object's category and location. The stated aim is to address a shortcoming in prior small-object detection accuracy.Filed by Chengdu Information Technology of CAS Co., Ltd., published 2023-04-20 as US20230122927A1.
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20190102646A1 | Image based object detection | 217 |
| 2 | US20180121762A1 | Neural network for object detection in images | 191 |
| 3 | US11205098B1 | Single-stage small-sample-object detection method based on decoupled metric | 99 |
| 4 | US10452959B1 | Multi-perspective detection of objects | 92 |
| 5 | JP2018077829A | Information processing method, information processing device and program | 88 |
| 6 | US20180300880A1 | Small object detection from a large image | 87 |
| 7 | US20190279046A1 | Neural network for object detection in images | 83 |
| 8 | US11631238B1 | Method for recognizing distribution network equipment based on raspberry pi multi-scale feature fusion | 71 |
| 9 | US20210089841A1 | Real-Time Object Detection Using Depth Sensors | 68 |
| 10 | US20210142097A1 | Image processing system | 67 |
Citation counts favour older filings that have had more time to accumulate references within this searched corpus — treat them as a signal of influence, not of current technical leadership.
Each row carries its publication number; clicking a row searches Eureka by that number.
Put your own technology through the same analysis
Eureka on the web
When you want the answer in the next five minutes.
The agent works the prompt against patents and technical literature, citing every source.
Run your analysis now →MCP server & REST API
When it has to run inside your own pipeline.
Patent search, landscape analysis and assignee resolution as MCP tools. Drop them into any agent framework, or call REST directly.
Browse MCP servers →What the numbers mean for a filing decision
Three patterns stand out once concentration, filing pace and technology composition are read together: no single assignee controls the space, growth has already plateaued, and claim density is heaviest where recognition and learning-model claims overlap.
No single assignee controls detection architecture claims
The leader holds 22 records and fifth place holds 13 — a real gap, but the top 5 combined still account for only 11.1% of all 733 records in scope. Even the top 10 combined reach just 18.6%. That leaves the large majority of filings spread across a long tail of single- or few-filing entrants.
Growth flattened after the early-2020s surge
Annual filings grew from 24 in 2017 to a peak of 112 in 2025, but the 2022 midpoint of 103 shows most of that growth was already banked by mid-period. The trend from 2022 onward is flat to declining rather than accelerating, which argues against treating this as a still-expanding filing category.
Recognition and AI-model classes co-occur on most filings
G06V (image/video recognition) sits on 73.0% of records and G06N (AI-model computing) on 52.4%, and since a record can carry multiple classes, most filings claim both together. G06T image processing follows at 42.6%, showing the pipeline from raw image data through recognition to a trained model is the dominant claim structure.
Medical, video and radar applications remain marginal but present
A61B (diagnosis/surgery), H04N (video communication) and G01S (radar/positioning) each sit near 3% of the 733 records. These are not core territory, but they show detection architectures are already being adapted into adjacent sensing and application domains at low but non-zero density.
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to object detection architectures, with the prior art for and against each one.
Who is filing, and where the field is still open
The ranked leaders span industrial electronics, mobile chipsets, cloud platforms and academic institutions rather than a single dominant filer. Recent-year momentum figures show even the most active recent filer slowing, which is consistent with the flat overall filing trend.
The top spot is well ahead of fifth place, but not the field
The leading assignee holds 22 records against 13 for fifth place and 10 for tenth — a genuine lead, but one that still represents a small fraction of the 733-record total. Recent-year momentum for this filer shows a -63% year-on-year drop, suggesting the lead was built earlier rather than through continuing acceleration.
Chipset and mobile platform assignees are pulling back, not pushing forward
One leading chipset assignee logged only 1 record in the latest year, down 50% year-on-year, while several other large technology-licensing assignees recorded zero filings in the latest year. That pattern across multiple large filers reinforces the flat-to-declining trend visible in the aggregate filing data.
Collaboration is limited and mostly bilateral
The corpus contains 10 identified co-assignee pairs, with the strongest pairings each appearing on 2 shared records. This is a modest amount of joint filing relative to 733 total records, indicating most assignees in this space file independently rather than through joint development programmes.
| Assignee | Recent year | YoY |
|---|---|---|
| VELLORE INSITUTE OF TECH | 6 | -63% |
| NVIDIA Corporation | 2 | — |
| Qualcomm Incorporated | 1 | -50% |
| Intel Corporation | 0 | — |
| Robert Bosch GmbH | 0 | — |
| Microsoft Technology Licensing, LLC | 0 | — |
| Telefonaktiebolaget LM Ericsson (publ) | 0 | -100% |
| Snap Inc. | 0 | — |
Where to take this analysis
The aggregate view answers who is filing and how densely; the open questions are which specific claims block a given design and where a first-to-file position is still available.
Check freedom-to-operate against the most-cited claims
The five most-cited records in this corpus date mostly from 2018-2019 and carry citation counts well above the rest of the set, making them the first place to check claim scope before committing to a detection architecture.
Explore the citation map in Eureka →Track the flattening filing trend by assignee
With aggregate filings flat since 2022 and several large assignees at zero filings in the latest year, tracking which specific companies are still active can reveal whether the plateau is field-wide or concentrated among a shrinking set of filers.
Set up assignee monitoring in Eureka →Test claim language against the under-claimed branches
Sub-areas like anchor-free small-object detection and latency-constrained edge inference show lighter filing density inside this corpus; drafting a first claim there benefits from checking exactly how existing claims are worded nearby.
Draft and compare claims in Eureka →Common questions about object detection architecture patents
Within this 733-record corpus, one assignee leads with 22 records, ahead of a fifth-place holder at 13 and a tenth-place holder at 10. However, the top 5 assignees combined hold only 11.1% of all 733 records, and the top 10 combined reach just 18.6%, so no single company controls a large share of the field. The remainder is spread across a long tail of the other ranked assignees, most with only a handful of filings each.
Filing activity grew substantially from 24 records in 2017 but had already reached 103 by the 2022 midpoint, close to the eventual peak of 112 in 2025. That pattern indicates the growth phase largely occurred in the earlier part of this period, with the trend flat to declining more recently rather than continuing to accelerate. The 2026 figure of 61 is a partial year and understated by publication lag, so it should not be read as a sudden drop.
G06V, covering image and video recognition, appears on 73.0% of the 733 records in scope, making it the dominant class. G06N, covering AI-model computing, follows at 52.4%, and because records can carry multiple IPC codes, most filings combine both — claiming a recognition pipeline together with a trained model rather than either alone. G06T image data processing appears on 42.6% of records, rounding out the core three-class pattern that defines most filings in this space.
Application-adjacent classes such as A61B (diagnosis and surgery), H04N (video communication) and G01S (radar and positioning) each account for only about 3% of the 733 records, showing detection architectures are being adapted into these domains but not yet heavily claimed there. Sub-areas like anchor-free small-object detection, label-noise-robust training and latency-constrained edge inference also show comparatively light claim density relative to the dominant recognition-plus-model pattern, making them worth a closer freedom-to-operate check before filing.
Citation counts in this corpus favour older filings — the most-cited records date mainly from 2018-2019 — because they have had more years to accumulate references from later filings. A high citation count signals historical influence within the searched corpus, not that the underlying architecture is still the most commercially relevant today. Recent-year momentum figures, which show several large assignees at zero filings in the latest year, are a better indicator of current activity than raw citation totals.
Research Object Detection Architectures in depth with Eureka
Go past this page: query the whole object detection architectures corpus yourself, in your own scope.
Every answer comes back with patent numbers you can open.
Disclaimer. This page is generated from Patsnap Eureka data drawn from a limited snapshot of global patent and scientific-literature records, and is provided for general information and reference only.
Patent data carries inherent limitations: recent filings (typically the most recent 18–24 months) are under-counted due to standard publication lag; counts may be reported at either a patent-family or a patent-record basis and are not always directly comparable; classification, applicant-name, and citation data may contain errors, duplicates, or omissions; and the underlying search query defines and constrains the scope shown. As a result, the analysis may be incomplete or inaccurate and may not reflect the full technology landscape.
Nothing on this page constitutes an exhaustive prior-art, novelty, freedom-to-operate, or validity search, nor does it constitute legal, financial, investment, or professional advice, and it should not be relied upon as such. Any patent, commercial, or strategic decision should be verified independently and reviewed with qualified patent, legal, and domain professionals. Patsnap makes no warranties, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of the information presented.
Machine translation. Assignee and organisation names originally recorded in Chinese, Japanese or Korean have been rendered into English by an AI translation step so that the tables stay readable. These renderings are best-effort and may not match a company’s registered English name; the original name is what the underlying patent record carries, and it is what any Eureka query launched from this page uses.