Efficient Person Search: An Anchor-Free Approach

Person search aims to simultaneously localize and identify a query person from uncropped images. To achieve this goal, state-of-the-art models typically add re-id branch upon two-stage detectors like Faster R-CNN. Owing the ROI-Align operation, pipeline yields promising accuracy as features are explicitly aligned with corresponding object regions, but in meantime, it introduces high computational overhead due dense anchors. In work, we present an anchor-free approach efficiently tackling challenging task, by introducing following dedicated designs. First, select detector (i.e., FCOS) prototype of our framework. Due lack anchors, exhibits significantly higher efficiency compared existing models. Second, when directly accommodating for search, there exist several misalignment issues different levels scale, region, task). address these issues, propose feature aggregation module generate more discriminative robust embeddings. Accordingly, name framework Feature-Aligned Search Network (AlignPS). Third, investigating advantages both anchor-based models, further augment AlignPS head, which improves robustness while still keeping model highly efficient. Our not only achieves or competitive performance on two benchmarks, can be also extended other searching tasks such animal search. All source codes, data, trained available at: https://github.com/daodaofr/alignps .