QA-ReID: Quality-Aware Dual-Branch Matching

• The paper introduces a dual-branch architecture that combines RGB appearance and clothing-invariant body-part features to overcome clothing-induced feature drift.
• It leverages a multi-modal attention fusion and pixel-level quality-aware similarity to enhance robustness in person re-identification.
• Benchmark results on PRCC, LTCC, and VC-Clothes datasets confirm state-of-the-art performance, demonstrating significant improvements over conventional methods.

Quality-Aware Dual-Branch Matching (QA-ReID) is a framework for person re-identification (ReID) under clothing change, designed to robustly distinguish identities in scenarios where appearance undergoes dramatic transformations due to varied garments. QA-ReID addresses the acknowledged limitations of conventional ReID, which are highly sensitive to clothing-induced feature drift, by leveraging a dual-branch architecture that extracts and fuses both appearance and clothing-invariant structural cues. At matching, a novel pixel-level quality-aware similarity is computed with adaptive weighting and geometric consistency constraints. Empirical results demonstrate consistent state-of-the-art performance across major clothes-changing person ReID (CC-ReID) benchmarks, substantially surpassing prior methods and conventional baselines (Wang et al., 27 Jan 2026).

1. Principles of QA-ReID: Problem Context and Motivations

Clothes-changing person re-identification (CC-ReID) introduces appearance shifts that fundamentally challenge the discriminative reliability of classic RGB-based ReID systems. Typical methods collapse when apparel is changed, as identity-relevant cues encoded in garments are no longer predictive. Key insight in QA-ReID is the joint use of: (i) global RGB appearance features and (ii) parsing-based, body-part-level features that are less affected by clothing variability. This hybrid representation is hypothesized to retain both coarse appearance context and fine-grained structural information, improving resistance to clothing changes. The quality-aware aspect further addresses intra-image region reliability by upweighting informative, invariant regions and suppressing garment-dependent artifacts (Wang et al., 27 Jan 2026).

2. Dual-Branch Feature Extraction and Representation

QA-ReID begins with two parallel pathways, each producing region-level feature maps from the input image $I\in\mathbb{R}^{3\times H'\times W'}$:

• Body Parsing and Masking: A human parser generates a dense semantic map $M = P(I)$ and a binary body mask $M_{\mathrm{body}}$. From this, a clothing-invariant version $I_{\mathrm{par}} = I\odot M_{\mathrm{body}}$ is derived, while the unmasked RGB image $I_{\mathrm{rgb}} = I$ is retained.
• Branch Processing: Both $I_{\mathrm{par}}$ and $I_{\mathrm{rgb}}$ are processed via identical ResNet-50 backbones up to stage 3, outputting $F_{\mathrm{par}}$, $$F_{\mathrm{rgb}} \in \mathbb{R}^{C\times H\times W}$$, respectively.

This two-stream approach ensures the extraction of robust identity cues grounded both in stable anatomical structures and variable appearance domains (Wang et al., 27 Jan 2026).

3. Multi-Modal Attention Fusion

To combine the heterogeneous modalities, QA-ReID employs a dual-attention module, with channel attention $A_c\in\mathbb{R}^{2C\times 1\times 1}$ and spatial attention $M = P(I)$0 computed on the concatenated feature map $M = P(I)$1. The final fused feature is given by:

$M = P(I)$2

$M = P(I)$3

$M = P(I)$4

where $M = P(I)$5 is broadcast and split into modality-specific attention masks. The fusion mechanism is designed to adaptively prioritize channels/regions most informative for identity under possible appearance shifts (Wang et al., 27 Jan 2026).

4. Quality-Aware Query-Adaptive Convolution (QAConv-QA)

For robust matching, QA-ReID introduces pixel-wise similarity computation with spatial quality assessment and bidirectional consistency:

• Pixel Quality Weights: Each pixel in $M = P(I)$6 receives a weight $M = P(I)$7 that reflects the proportion of 'body' pixels in its associated $M = P(I)$8 patch. This gives higher influence to spatial locations over body parts less impacted by garment (e.g., head, limbs).
• Weighted Cosine Similarity: The similarity between two feature pixels incorporates both quality weights and cosine similarity, i.e., $M = P(I)$9.
• Bidirectional Consistency: Similarities are normalized in both directions (probe-to-gallery and gallery-to-probe) and multiplied: $M_{\mathrm{body}}$0. The resulting map is pooled with Bi-GMP (Bidirectional Global Max Pool) to yield an image-level similarity $M_{\mathrm{body}}$1, which is passed through an MLP and sigmoid to obtain a matching probability $M_{\mathrm{body}}$2.

This module operationalizes region reliability and mutual salience in the identity comparison (Wang et al., 27 Jan 2026).

5. Loss Functions and End-to-End Optimization

QA-ReID is trained with a composite loss:

• Branch Classification and Triplet Losses: Softmax-based classification and batch-hard triplet loss are applied independently on the pooled features of both $M_{\mathrm{body}}$3 and $M_{\mathrm{body}}$4.
• Pairwise Matching Loss: A binary cross-entropy loss is imposed on the predicted pairwise matching probabilities $M_{\mathrm{body}}$5 for all pairs within each batch.

The aggregate loss is:

$M_{\mathrm{body}}$6

This encourages both discriminative branch learning and directly optimizes for cross-appearance matching (Wang et al., 27 Jan 2026).

6. Algorithmic Workflow and Pseudocode

The full QA-ReID training procedure comprises repeated minibatch processing, including parsing, masking, dual-branch feature extraction, adaptive fusion, computation of classification/triplet/matching losses, and parameter updates. The key operational steps are illustrated in the following pseudocode (verbatim, all notation as in the source) (Wang et al., 27 Jan 2026):

$M_{\mathrm{body}}$7 (Wang et al., 27 Jan 2026)

7. Benchmark Results and Comparative Assessment

The efficacy of QA-ReID is established through systematic evaluation on PRCC, LTCC, and VC-Clothes datasets under both same-clothing and clothing-changing conditions. In direct comparison with conventional methods (PCB, TransReID) and prior CC-ReID approaches (notably CLIP3DReID and MCSC), QA-ReID exhibits:

• PRCC: 64.1% Top-1, 61.2% mAP under changing-clothes protocol (up to +5% improvement over the previous best).
• LTCC: 42.9% Top-1, 21.3% mAP (comparable to state-of-the-art).
• VC-Clothes: 86.3% Top-1, 86.1% mAP (over +3 mAP improvement).

Consistently, conventional methods display severe performance degradation under clothing changes (PCB: 38.7 mAP; TransReID: 49.3 mAP on PRCC), while QA-ReID achieves clear state-of-the-art results, supporting the hypothesis that adaptive fusion of appearance and body structure—together with quality-aware pixel-level matching—substantially enhances robustness to garment variability (Wang et al., 27 Jan 2026).

Dataset	Protocol	QA-ReID Top-1 (%)	QA-ReID mAP (%)	Best Prior (mAP)
PRCC	Clothes-Changing	64.1	61.2	59.3 (CLIP3DReID)
LTCC	Clothes-Changing	42.9	21.3	21.7 (CLIP3DReID)
VC-Clothes	Clothes-Changing	86.3	86.1	83.2 (MCSC)

All results are as reported in (Wang et al., 27 Jan 2026).

---

Quality-Aware Dual-Branch Matching (QA-ReID) extends the principle of dual-branch feature extraction from the set-to-set recognition literature, notably the Quality Aware Network (QAN) (Liu et al., 2017), to the image-to-image CC-ReID problem, but goes further by introducing adaptive multi-modal fusion and highly local, quality-weighted similarity that handles region-specific reliability and bidirectional identity consistency. In aggregate, these innovations advance the robustness and discrimination capacity of person re-identification under challenging appearance changes due to clothing, as corroborated by comprehensive benchmark evaluations.