This dataset tracks 647 patent families published between 2015 and the 2026 data cut-off, scoped to filings whose claims or descriptions cover encoder-decoder architectures, receptive field design, boundary refinement, annotation cost, or domain adaptation, within the G06V10, G06T7 and G06V20 IPC classes. That figure counts families, which is a fairer measure than raw document counts because it neutralises repeat continuation filings and multi-jurisdiction duplicates of the same invention. Because publication lags filing by roughly 18 months, the true count for 2024-2026 filings will continue to rise as more records publish.

China is the leading receiving office in this dataset with 304 records, ahead of the United States at 195, and well ahead of Europe, WIPO/PCT, Australia and India, which each account for a small fraction of the total. This matters for competitive tracking and clearance work alike: a review that only checks USPTO and EPO filings is screening against the second-largest pool, not the largest. Firms operating or selling into China should treat Chinese-language filings as a primary rather than supplementary search target.

No — filings rose from 21 in 2017 to a peak of 144 in 2023, then declined, with the 2022 midpoint at 102 suggesting the growth phase had already begun flattening before the peak. The drop shown for 2025 and 2026 is partly an artefact of the roughly 18-month gap between filing and publication, so the most recent two years will fill in somewhat as more records publish. Even accounting for that lag, the overall shape is a matured filing wave rather than an accelerating one, consistent with core architectural claims having been staked out earlier in the window.

The dataset's assignee ranking spans global electronics and AI firms, autonomous-vehicle specialists, and Chinese universities, with filing density concentrated in a relatively small number of names against a long tail of single-filing entrants. Notably, the six assignees showing the strongest recent-year momentum in aggregate recorded zero filings in the most recent tracked year, which — allowing for publication lag — suggests consolidation of existing claim positions rather than active expansion at the moment. Co-filing is uncommon: only eight co-assignee pairs appear across all 647 families, with the strongest recurring just three times.

Filing density is heaviest around core encoder-decoder architectures and boundary-refinement methods, classified primarily under G06V and G06T. Thinner coverage appears in cross-sensor domain adaptation, point-cloud instance boundary refinement, annotation-cost reduction through pseudo-labelling, and remote-sensing-specific cross-domain segmentation — branches that show up in the dataset but with fewer dedicated filings than the core architectural claims. High density in the core does not mean those methods are settled science, only that the claim space there is already occupied; the adjacent branches are where a first claim is more likely to clear without heavy prior art.