DC-VLAQ: Robust VPR & Dynamic VLC Allocation

• DC-VLAQ is a dual-framework integrating query-residual aggregation to create robust global descriptors by fusing DINOv2 and CLIP features for visual place recognition.
• The method utilizes VLAQ pooling to aggregate local tokens, effectively preserving fine-grained spatial cues and achieving state-of-the-art Recall@1 on multiple benchmarks.
• In VLC, DC-VLAQ dynamically allocates optical channels based on real-time demand, ensuring high-priority QoS while maintaining overall channel utilization.

DC-VLAQ denotes two distinct advanced frameworks in research literature: (1) "Query-Residual Aggregation for Robust Visual Place Recognition" for constructing domain-robust global visual descriptors, and (2) "Dynamic Channel Allocation for QoS Provisioning in Visible Light Communication" as a real-time differentiated-service resource allocation protocol. Each instance serves as a state-of-the-art solution in its respective domain, distinguished by residual-based fusion or resource reservation mechanisms.

1. DC-VLAQ in Visual Place Recognition: Representation-Centric Fusion and Query-Residual Aggregation

In visual place recognition (VPR), DC-VLAQ is a representation-centric pipeline that addresses the challenge of constructing robust global image descriptors resilient to large viewpoint variation, illumination change, and significant domain shifts. The core innovation lies in the integration of complementary Visual Foundation Models (VFMs) using a residual-guided fusion strategy, anchored on the DINOv2 feature space with residual semantic enrichment from CLIP. This is coupled with a query-residual aggregation mechanism—Vector of Local Aggregated Queries (VLAQ)—that encodes local tokens by their deviations from learnable query vectors, thus stabilizing pooling under distribution shifts and preserving fine-grained cues (Zhu et al., 19 Jan 2026).

2. Residual-Guided Complementary Fusion

The fusion module combines two sets of local token features for each image:

• $X_i^D \in \mathbb{R}^{M \times d}$: DINOv2 tokens (appearance-anchored)
• $X_i^C \in \mathbb{R}^{M \times d}$: CLIP tokens (semantically enhanced)

Fusion is formalized as:

$z_{ij} = x^D_{ij} + F_C(x^C_{ij} - x^D_{ij})$

where $F_C: \mathbb{R}^d \to \mathbb{R}^d$ is a learned linear layer, scaling/rotating the CLIP-to-DINOv2 residual for each token $j$. L2 normalization is applied to the original tokens before fusion. Only the terminal two DINOv2 blocks are fine-tuned, with the CLIP encoder remaining frozen. This preserves the DINOv2 geometric anchor while leveraging CLIP's complementary semantics, avoiding conflicts in embedding space and supporting stable downstream aggregation.

3. Vector of Local Aggregated Queries (VLAQ) Aggregation

Global descriptor formation employs the Vector of Local Aggregated Queries, a residual-to-learnable-query pooling that generalizes the evolution from Bag-of-Words (BoW) to VLAD, mitigating instability induced by multi-backbone fusion.

For $S$ learnable queries $\{q_k\}_{k=1}^S$, for each token $z_{ij}$:

• Compute scaled dot-product scores:

$s_{ijk} = \frac{q_k^\top z_{ij}}{\sqrt{d}}$

• Soft-assign tokens to queries:

$$\alpha_{ijk} = \frac{\exp(s_{ijk})}{\sum_{j'=1}^M \exp(s_{ij'k})}$$

• Encode residual response:

$v_{ik} = \sum_{j=1}^M \alpha_{ijk} (z_{ij} - q_k)$

The global descriptor is then

$$\bar{g}_i = [v_{i1}^\top, ..., v_{iS}^\top]^\top, \quad g_i = \frac{\bar{g}_i}{\|\bar{g}_i\|_2}$$

This approach ensures insensitivity to absolute magnitude and distribution shifts, while retaining fine-grained spatial and semantic discrimination.

4. End-to-End DC-VLAQ Visual Pipeline

Key algorithmic steps:

Stage	Operation	Output Dimension
Image Preprocessing	Resize $I_i$ to $280 \times 280$ (train) or $322 \times 322$ (test)	–
Local Feature Extraction	$X_i^D = \mathcal{E}_{DINO}(I_i)$, $X_i^C = \mathcal{E}_{CLIP}(I_i)$	$M \times d$
Residual Fusion	$Z_i = X_i^D + F_C(X_i^C - X_i^D)$	$M \times d$
VLAQ Aggregation	S=64 queries, B=2 blocks (multi-block), residual pooling	$B \cdot S \cdot d$
Descriptor Normalization	L2 normalize output descriptor	$49,\!152$ (e.g., 2×64×384)

Training uses GSV-Cities dataset, Multi-Similarity loss, and AdamW optimizer. Evaluation uses Recall@K on benchmarks (Pitts30k, Tokyo24/7, MSLS, Nordland, SPED, AmsterTime).

5. Quantitative Evaluation and Comparative Performance

On standard VPR benchmarks, DC-VLAQ demonstrates consistent and often state-of-the-art Recall@1 across diverse datasets and challenging conditions:

Benchmark	BoQ Recall@1 (%)	DC-VLAQ Recall@1 (%)
Pitts30k-test	93.7	94.3
Tokyo24/7	98.1	98.7
MSLS-val	93.8	94.2
MSLS-challenge	79.0	81.7
Nordland	90.6	92.8
SPED	92.5	93.9
AmsterTime	63.0	66.8

These results reflect superior stability and fine-grained retrieval, especially under substantial domain shift and temporal variation, consistently outperforming baseline methods including BoQ, NetVLAD, SFRS, MixVPR, and others (Zhu et al., 19 Jan 2026).

6. DC-VLAQ for Dynamic Channel Allocation in Visible Light Communication

DC-VLAQ also refers to "Dynamic Channel Allocation for QoS Provisioning in Visible Light Communication" (Chowdhury et al., 2018). This scheme dynamically reserves optical (color) channels for higher-priority traffic classes in Visible Light Communication (VLC) systems based on real-time Poisson arrival rate estimates. It's designed to optimize both blocking probability (favoring high-priority traffic) and overall channel utilization without sacrificing system throughput.

Key model elements:

• $N$: total available channels
• $M$: priority classes (with $m=1$ highest)
• Dynamic thresholding based on instantaneous estimated arrival rates $\hat\lambda_m$
• Guard pool $G=N-C$, with per-class allocation $R_m=\frac{\hat\lambda_m}{\hat\lambda_T}G$
• Class-$m$ admitted if busy channels $<N_m = C+\sum_{k=1}^{m-1}R_k$

Analytically, blocking and utilization are characterized by the M/M/N/N occupancy:

$$B_m = \sum_{n=N_m}^N P(n), \quad U = \frac{1}{N} \sum_{n=0}^N nP(n)$$

with $P(n)$ as usual equilibrium probabilities. The approach achieves less than 1% blocking for highest-priority calls and above 80% utilization across loading conditions, outperforming non-priority static sharing without capacity loss (Chowdhury et al., 2018).

7. Impact and Significance

In VPR, the DC-VLAQ paradigm demonstrates that anchoring fusion on an appearance-focused model with residual semantic enrichment, coupled with query-residual aggregation, leads to robust, stable, and discriminative global representations. These innovations enable strong performance under severe domain shifts, long-term environmental change, and diverse benchmarking scenarios.

For VLC resource allocation, the DC-VLAQ framework delivers differentiated quality of service (QoS) by real-time guard-channel reservation proportional to observed demand. This dynamic allocation sharpens prioritize for delay-sensitive traffic while retaining high channel occupancy, a key requirement for high-performance, mixed-service wireless networks.

Both instances of DC-VLAQ illustrate that residual-centric aggregation—whether for feature fusion or resource allocation—can yield state-of-the-art robustness and efficiency in the face of noisy or multi-modal input distributions, setting clear baselines in their respective fields (Zhu et al., 19 Jan 2026, Chowdhury et al., 2018).