QAConv-QA: Quality-Aware Query-Adaptive Conv

• The paper introduces QAConv-QA, a pixel-level matching mechanism that adaptively generates convolution filters and enforces bidirectional constraints to robustly address clothes-changing re-identification.
• It integrates dual-branch feature extraction with multi-modal attention fusion to combine RGB and parsing cues, producing discriminative and clothing-invariant representations.
• Experiments on PRCC and LTCC benchmarks show significant performance gains, confirming the effectiveness of spatial quality weighting and bidirectional matching in noisy environments.

Quality-Aware Query-Adaptive Convolution (QAConv-QA) is a pixel-level matching framework developed to address the problem of clothes-changing person re-identification (CC-ReID), where traditional appearance cues become unreliable due to significant intra-identity clothing variations. QAConv-QA extends earlier query-adaptive convolution methods by incorporating spatially-varying quality weighting and enforcing bidirectional matching constraints, thereby enhancing robustness to clothing variation and background noise. The module is integrated into the broader QA-ReID architecture, which employs dual-branch feature extraction and multi-modal attention fusion to jointly leverage global appearance and clothing-invariant structural cues (Wang et al., 27 Jan 2026).

1. Integration within Dual-Branch Feature Extraction

QAConv-QA operates on fused feature maps within QA-ReID. Each image $I \in \mathbb{R}^{3 \times H' \times W'}$ is passed through two parallel ResNet-50 branches truncated at stage-3: an RGB branch producing $$F_{\text{rgb}} \in \mathbb{R}^{C \times H \times W}$$, and a parsing branch generating a binary body mask $M_{\text{body}} \in \{0,1\}^{H'\times W'}$ and extracting $$F_{\text{par}} \in \mathbb{R}^{C \times H \times W}$$. A multi-modal attention module computes channel- and spatial-wise attention weights $\omega \in [0,1]^{C \times H \times W}$ across the concatenated features $[F_{\text{rgb}}; F_{\text{par}}]$. The fused features are obtained as:

• $$F_{\text{mix}} = \omega \odot F_{\text{rgb}} + (1-\omega) \odot F_{\text{par}}$$
• $$F = F_{\text{rgb}} + F_{\text{par}} + F_{\text{mix}}$$
• $F_{\text{fuse}} =$ 1×1 convolution($F$), resulting in $$F_{\text{fuse}} \in \mathbb{R}^{C \times H \times W}$$.

QAConv-QA is applied to pairs of these fused features from probe and gallery images to compute a fine-grained similarity score.

2. Query-Adaptive Convolution Filter Generation

QAConv-QA dynamically generates a local convolution kernel at each spatial position of the probe's fused feature map. Specifically, for the probe $F_q = F_{\text{fuse}}$ of dimension $C \times H \times W$, a learnable 1×1 convolution $\phi : \mathbb{R}^C \rightarrow \mathbb{R}^{C k^2}$ is applied to yield $$W_q = \phi(F_q) \in \mathbb{R}^{C k^2 \times H \times W}$$. The slice at each $(i, j)$ is reshaped into a kernel $w_q(i, j) \in \mathbb{R}^{C \times k \times k}$, which is convolved over the gallery feature $F_g \in \mathbb{R}^{C \times H \times W}$ at the corresponding position. Stacking these responses for all $(i, j)$ results in a raw pixel-to-pixel similarity tensor.

This procedure allows for localized, dynamic, query-conditioned matching, in contrast to global or fixed-filter approaches.

3. Pixel-Level Importance Weighting

Not all spatial locations in the feature map contribute equally to identity discrimination, especially under clothing changes and background clutter. QAConv-QA computes a spatial quality (or importance) map $Q \in \mathbb{R}^{H \times W}$, derived from the downsampled body mask:

• The original mask $M_{\text{body}}$ is pooled over $k \times k$ patches:

$$\tilde{Q}_{i,j} = \frac{1}{k^2} \sum_{u=i k}^{i k + k-1} \sum_{v=j k}^{j k + k-1} M_{\text{body}}(u, v)$$

• A spatial softmax normalizes these coarse mask sums:

$$Q_{i,j} = \frac{\exp(\tilde{Q}_{i,j})}{\sum_{p=1}^H \sum_{q=1}^W \exp(\tilde{Q}_{p,q})}$$

During matching, the score for a pixel pair $(f^1_{i_1, j_1}, f^2_{i_2, j_2})$ is weighted as follows:

$$\text{sim}^1(f^1_{i_1, j_1}, f^2_{i_2, j_2}) = Q^1_{i_1, j_1} \cdot Q^2_{i_2, j_2} \cdot \rho(f^1_{i_1, j_1}, f^2_{i_2, j_2})$$

where $\rho(\cdot, \cdot)$ denotes cosine similarity.

This weighting suppresses contributions from background and unreliable regions, emphasizing body structure over appearance.

4. Bidirectional Consistency Constraints

Single-directional similarity matching can yield high scores for spurious correspondences. QAConv-QA introduces bidirectional normalization to enforce mutual matching. For each pair $(f^1_{i_1, j_1}, f^2_{i_2, j_2})$:

• Forward-normalized score:

$$\bar{\rho}(f^1_{i_1, j_1} | f^2_{i_2, j_2}) = \frac{\exp(\text{sim}^1(f^1_{i_1, j_1}, f^2_{i_2, j_2}))}{\sum_{h=1}^H \sum_{w=1}^W \exp(\text{sim}^1(f^1_{h, w}, f^2_{i_2, j_2}))}$$

• Reverse-normalized and combined:

$$\text{sim}^2(f^1_{i_1, j_1}, f^2_{i_2, j_2}) = \bar{\rho}(f^1_{i_1, j_1} | f^2_{i_2, j_2}) \times \bar{\rho}(f^2_{i_2, j_2} | f^1_{i_1, j_1})$$

Only pairs that are the mutually most likely correspondences yield large contributions. A bidirectional global max-pooling (Bi-GMP) operation further aggregates the similarity map to a scalar, followed by batch normalization, an MLP, and sigmoid to produce a final match probability $p \in [0, 1]$.

This mechanism explicitly penalizes one-way mismatches, increasing reliability under appearance changes.

5. Matching Workflow

A probe-gallery matching instance proceeds as follows:

1. Extract $F_{\text{rgb}}$, $F_{\text{par}}$ for both probe and gallery via respective branches.
2. Fuse with attention $\omega$ to generate $F_{\text{fuse}}^q$, $F_{\text{fuse}}^g$.
3. Generate $H \times W$ query-adaptive filters $w_q(i, j)$ from $F_{\text{fuse}}^q$.
4. Convolve each $w_q(i, j)$ over $F_{\text{fuse}}^g$ to obtain raw responses.
5. Compute pixel-level quality maps $Q^q$, $Q^g$ and weight the responses.
6. Apply bidirectional normalization as above.
7. Aggregate with Bi-GMP, apply batch normalization, MLP, and sigmoid to yield the scalar similarity $p_{q, g}$.

6. Training Objectives and Loss Formulation

The QA-ReID framework containing QAConv-QA is trained end-to-end using three loss components:

• Cross-entropy identity losses on global pooled features of each branch:

$$L_\text{cls}^{\text{rgb}} + L_\text{cls}^{\text{par}}$$

• Triplet losses on both branches:

$$L_\text{tri}^{\text{rgb}} + L_\text{tri}^{\text{par}}$$

• Pairwise matching loss for QAConv-QA:

$$L_\text{match} = -\frac{1}{B^2} \sum_{i=1}^B \sum_{j=1}^B [y_{ij} \log p_{ij} + (1-y_{ij}) \log (1-p_{ij})]$$

where $y_{ij} = 1$ if images $i$ and $j$ share the same ID.

The overall loss is:

$$L = L_\text{cls}^{\text{rgb}} + L_\text{cls}^{\text{par}} + L_\text{tri}^{\text{rgb}} + L_\text{tri}^{\text{par}} + L_\text{match}$$

This joint objective encourages discriminative, identity-preserving representation in both color/texture and structure, and robust pixel-level matching under clothing changes.

7. Performance Analysis and Impact

Empirical results substantiate the effectiveness of QAConv-QA’s mechanisms in the CC-ReID setting. Experiments on PRCC and LTCC benchmarks show:

• Baseline fusion with plain QAConv (no pixel weighting, no bidirectional constraints) attains 61.4% Top-1 on PRCC and 41.1% on LTCC.
• Adding pixel weighting alone increases performance to 63.0% / 41.8%.
• Incorporating bidirectional matching singly yields 63.7% / 42.4% (value for LTCC inferred by juxtaposition, this suggests bidirectional constraints alone moderately help).
• The full QAConv-QA with both enhancements achieves 64.1% Top-1 on PRCC and 42.9% on LTCC, representing a $+2.7\%$ absolute gain on PRCC over QAConv and $+1.8\%$ over the strongest existing competitor.

The explicit use of spatial quality weighting and reciprocal normalization distinguishes QAConv-QA from conventional global-matchers and the original QAConv, leading to increased robustness against large clothing changes, background interference, and unreliable regions (Wang et al., 27 Jan 2026).