Anchor-Free Person Search

1. Introduction

• Two-step approaches separate detection and re-id, whereas one-step models unify them but commonly use two-stage detectors with high computational complexity and anchor-related hyperparameter sensitivity.The one-step methods described apply ROI-Align to detected boxes and share those features between detection and re-id.Source: §1 · pp. 1–2
• Existing anchor-free detectors pose scale, region, and task misalignment challenges for person-search re-id.Multi-scale features complicate matching a query against gallery people at varied scales; without ROI-Align, re-id embeddings must be learned directly from feature maps, while detection and re-id require a suitable task balance.Source: §1 · p. 2
• AlignPS is the first anchor-free framework for efficient person search, using an aligned feature aggregation module and a “re-id first” principle.The module uses deformable convolution and feature fusion to address scale and region misalignment, while training gives more emphasis to robust re-id embeddings.Source: §1 · p. 2
• The paper presents an anchor-free one-step, one-stage framework for efficient person search.The proposed approach is intended to foster future research in this direction.Source: §1 · p. 2
• AlignPS outperforms state-of-the-art one-step two-stage models on two challenging person-search benchmarks while running at higher speed.The authors also report that it outperforms anchor-based competitors.Source: §1 · p. 2

2. Related Work

• One-stage anchor-free detectors are attracting attention for their simple structures and efficient implementations [36] [23] [56] [50] [42].AlignPS develops its person-search framework from a classic one-stage anchor-free detector.Source: §2 · p. 2
• Person re-id learns identity embeddings from cropped images and requires fine-grained identity features, unlike detection’s common pedestrian features.Prior re-id work includes body-structure and part information [40] [41] [48], distance-metric learning [20] [8] [9] [44], and data augmentation [16] [45] [55].Source: §2 · p. 2
• Person-search methods are divided into two-step pipelines and one-step joint models, with one-step methods generally simpler and more efficient but with room to improve their detector limitations.Two-step approaches detect and crop people before re-id; one-step methods commonly build on two-stage anchor-based detectors.Source: §2 · pp. 2–3

3.1. Framework Overview

• AFA output features serve directly as re-id embeddings, while an FCOS detection head predicts box offsets, centerness, and foreground/background classification.Each AFA output location is associated with a box, classification and centerness scores, and a re-id embedding.Source: §3.1 · p. 3

3.2. Aligned Feature Aggregation

• Using only P3 for detection and re-id addresses scale misalignment, with a possible slight effect on detection performance.P3 is the largest AFA output feature map; the model does not generate P6 or P7.Source: §3.2 · pp. 3–4
• Deformable convolutions in the lateral and output layers, plus concatenation in the top-down pathway, address region misalignment.The deformable convolutions adapt receptive fields, while concatenation replaces FPN’s sum operation to aggregate multi-level features.Source: §3.2 · p. 4
• The re-id-first design directly supervises AFA output features before they enter the detection head, avoiding an extra re-id embedding layer.The authors report improved embedding discriminability with negligible detection impact, and greater efficiency than detection-first or parallel structures.Source: §3.2 · p. 4

3.3. Triplet-Aided Online Instance Matching Loss

• TOIM adds triplet comparisons to OIM, whose distances omit input-to-input comparisons and whose log-likelihood lacks an explicit pairwise distance metric.It samples features around person centers and from the lookup table to form positive and negative pairs, then combines triplet and OIM losses.Source: Figure 3 · p. 4, §3.3 · p. 5

4.1. Datasets and Settings

• Experiments evaluate person search on CUHK-SYSU and PRW using mAP and top-1 accuracy, and detection using recall and AP; CUHK-SYSU uses the standard training/test split.Inference images are rescaled to 1500×900; training uses multi-scale resizing and an ImageNet-pretrained ResNet-50 backbone.Source: §4.1 · p. 5, §4.2 · p. 5

4.3. Analytical Results

• On the FCOS baseline with a re-id head parallel to detection, combining scale, region, and task alignment yields >20% improvement in mAP; each strategy also improves the baseline.Assigning scales to feature levels with the reported ranges slightly improves detection recall, but brings back re-id scale misalignment; multi-scale training does not well address it.Source: §4.3 · pp. 5–6, Figure 4 · p. 5
• Without region-alignment modules, top-1 accuracy is 83.7%; the full combination reaches 93.4% top-1 accuracy and 93.1% mAP.Lateral and output deformable convolutions each improve performance, while feature concatenation adds a smaller gain; C3 offsets are tighter than complementary, broader C4 offsets.Source: §4.3 · p. 6, Table 2 · p. 6, Figure 5 · p. 6
• The re-id-first task structure outperforms shared-head and independent-head alternatives for aligning detection and re-id.T1 and T2 share re-id features with detection heads and yield lower re-id performance; independent-head T3 performs slightly better than T1 and T2 but remains below the proposed design.Source: Figure 6 · p. 6, §4.3 · p. 7, Table 3 · p. 7
• TOIM improves mAP and top-1 accuracy by 0.7% and 0.5%, respectively; backbone deformable convolutions add only ∼1%, less than AFA layers.The TOIM gain comes from considering relations between input features and features stored in the LUT.Source: §4.3 · p. 7, Table 4 · p. 7, Table 5 · p. 7

4.4. Comparison to the State-of-the-Arts

• On CUHK-SYSU, AlignPS+ beats NAE+ by 1.9% in mAP and 1.6% in top-1 accuracy; AlignPS outperforms all one-step models but trails TCTS.The gallery-size comparison also shows AlignPS outperforming all one-step models and only falling short of the strongest two-step model, TCTS.Source: §4.4 · pp. 7–8, Table 6 · p. 7, Figure 8 · p. 8
• On PRW, AlignPS outperforms all one-step methods, while BINet also performs strongly but requires an additional re-id branch for region alignment.PRW has less training data, and all models perform worse on it than on CUHK-SYSU.Source: Table 6 · p. 7, §4.4 · p. 8
• 61 milliseconds per image makes AlignPS 27% faster than NAE and 38% faster than NAE+ in the same-GPU comparison.The comparison used a Tesla V100 GPU and images resized to 1500×900; AlignPS computes gallery features once.Source: §4.4 · p. 8, Table 7 · p. 8

• AlignPS outperforms existing person search methods and runs at a higher speed.Source: §5 · p. 8
• Its aligned feature aggregation module addresses scale, region, and task misalignment when adapting an anchor-free detector to person search.Source: §5 · p. 8
• The framework jointly addresses detection and re-id in a one-step anchor-free model.Source: §5 · p. 8