Writer Retrieval at Scale

Writer retrieval benchmarks usually assume a full page of connected handwriting. Real archives more often look like FormWR: 387,000 pages of printed forms carrying only a few handwritten fields each.

My contribution is X-VLAD, the learned set-aggregation module that collapses a variable number of local descriptors into one compact writer embedding. Transferred to the HisFragIR20 competition benchmark, the method reaches 78.8% mAP and 97.9% Top-1 — an improvement of +21.6% mAP and +12.7% Top-1 over prior published work, from a 256-dimensional embedding.