PaperlayerView exact revision on arXiv

Source-linked AI summary

Adaptive Semantic Communication with Residual Quantization for Resilient Vehicular Networks

Xuanhao Luo, Ruichen Gao, Sandeep Sreenivasan, Zhizhen Li, Yuchen Liu

arXiv:2610.10886v1cs.NI

TL;DR

Bandwidth-limited, changing vehicular links make it difficult to deliver task-relevant visual data reliably and on time, while conventional codecs do not account for semantic relevance. The paper combines residual quantization and context-adaptive lossless coding with runtime selection of quantization depth. Its system-level evaluation reports improved delivery reliability and timely transmission versus traditional image transmission, alongside low-latency operation and robust object detection.

  • Problem

    Conventional pixel-level codecs allocate bits without regard to downstream task semantics, while dynamic vehicular links make efficient, timely visual-data delivery challenging.

    Source: §I · p. 1

  • Method

    RQ-NAC combines residual quantization and context-adaptive lossless entropy coding, with runtime depth adaptation to adjust the transmitted representation to communication conditions.

    Source: §I · p. 2, §IV · p. 4

  • Results

    System-level evaluation reports improved delivery reliability and timely transmission versus traditional image transmission, with low-latency operation and robust object detection performance.

    Source: §VIII · p. 13

  • Takeaways & Limitations

    The evaluation jointly considers perception-task accuracy, delivery success, and latency to assess whether transmitted visual information remains useful under communication and computing constraints.

    Source: §I · p. 2

Abstract

from arXiv · show

Vehicular networks require timely visual information exchange to support safety-critical applications such as cooperative perception and hazard awareness, yet transmitting high-dimensional camera data over bandwidth-limited and dynamic wireless links remains challenging. In this paper, we propose an adaptive semantic communication architecture for vehicular multicast that adjusts the amount of visual information transmitted according to current channel conditions, available communication resources, and latency requirements. Specifically, the visual encoder uses residual quantization to represent each image at multiple bitrate and reconstruction-quality levels. At runtime, the system selects an appropriate level based on the communication conditions, enabling it to trade visual fidelity for lower transmission overhead when network resources are limited. Context-adaptive entropy coding further compresses the quantized representations using offline-learned statistics, without requiring retraining for different channel conditions. This design allows the visual representation to be learned offline while the communication strategy is adapted online. Experimental results demonstrate competitive rate-distortion performance at ultra-low bitrates, improved delivery reliability and deadline satisfaction across diverse vehicular scenarios, and strong downstream perception performance, achieving up to 0.88 mAP50 while dynamically adapting to changing communication conditions.

I. INTRODUCTION

Dynamic vehicular links make efficient, reliable delivery of task-relevant visual data difficult, while conventional codecs ignore scene semantics and channel-coupled methods can depend on training conditions. The paper proposes scalable RQ-NAC compression with runtime depth adaptation and extends prior semantic-compression work toward system-level vehicular evaluation.

I. INTRODUCTION

  • Existing approaches leave a system-level gap: channel-coupled JSCC may require adaptation or retraining outside its training regime, while generative methods can produce semantically inconsistent reconstructions.The paper identifies latency, delivery reliability, downstream perception utility, and runtime adaptation as joint concerns for dynamic vehicular networks.Source: §I · p. 1, §II · p. 3
  • Runtime-selectable pretrained residual-quantization configurations adapt transmitted representations to changing channel, resource, and latency conditions without online retraining.The controller jointly considers effective multicast capacity, receiver scale, and application deadlines to trade reconstruction fidelity against timely delivery.Source: §I · p. 2
  • RQ-NAC combines residual quantization for controllable bitrate-fidelity operating points with context-adaptive lossless entropy coding of residual index streams.Its representation learning is channel-agnostic, while lightweight depth adaptation makes runtime communication channel-aware.Source: §I · p. 2

II. RELATED WORK

  • Unlike channel-tailored JSCC and generative SemCom, this work targets scalable ultra-low-bitrate visual source coding while adapting residual-quantization depth to communication and latency constraints.The paper contrasts this with discrete VQ approaches that still rely on single-stage quantization, which it describes as limiting representational capacity and scalability at constrained bitrates.Source: §II · pp. 2–3
  • This paper extends the authors’ preliminary RQ-NAC framework with runtime depth adaptation, multicast and latency constraints, block-level HARQ modeling, and downstream perception analysis.The stated shift is from standalone semantic compression to an adaptive system for dynamic vehicular networks.Source: §II · p. 3

III. SYSTEM MODEL

The system sends compact visual representations from a source vehicle to multiple receivers and selects the residual-quantization operating point using channel and system conditions. Its communication model accounts for heterogeneous multicast receivers, shared resource allocation, and block retransmissions when estimating delivery time.

III. SYSTEM MODEL

  • The transmitter selects an RQ operating point from observed channel state and system constraints, trading compression efficiency, latency, and reconstruction quality for multicast delivery.Receivers decode the bitstream and reconstruct visual content for downstream perception.Source: §III · p. 3

A. Communication Model

  • The model represents multicast channel quality by the (1 − η)-quantile of receiver SINR and determines effective multicast rate from that value and each flow’s allocated resources.The per-flow resource fraction depends on the total visual-dissemination fraction and the number of concurrent flows.Source: A. Communication Model · pp. 3–4

B. Block Transmission and HARQ Model

  • Compressed bitstreams are segmented into independently decoded transport blocks, with a simplified HARQ model allowing each block up to Nretx retransmissions after failure.The resulting block transmission latency is included in the runtime depth-selection model.Source: B. Block Transmission and HARQ Model · p. 4

C. Representative Operating Scenarios

The paper evaluates representative vehicular conditions and presents RQ-NAC as a two-stage pipeline for scalable semantic compression. Its residual quantizer produces hierarchical discrete representations whose depth controls the transmitted representation.

C. Representative Operating Scenarios

  • Cruise, Junction, and Collision represent operating conditions ranging from relatively relaxed to highly constrained, differing in multicast scale, channel, load, and latency requirements.Detailed parameter settings are given in the evaluation methodology.Source: C. Representative Operating Scenarios · p. 4

IV. PROPOSED TWO-STAGE SEMANTIC COMPRESSION FRAMEWORK

  • RQ-NAC combines a residual-quantization semantic encoder with a lossless context-adaptive entropy-coding module in a two-stage compression pipeline.The pipeline produces a scalable representation for efficient transmission.Source: §IV · p. 4, Figure 2 · p. 5

A. Residual Quantization for Semantic Compression

  • The encoder maps visual input to hierarchical discrete latent representations with controllable depth, forming a code map that the receiver uses to reconstruct the latent map and image.The index sequence is formed from the quantization code map and passed to the entropy-coding stage.Source: A. Residual Quantization for Semantic… · pp. 4–5

B. Context-adaptive Arithmetic Coding

Context-adaptive arithmetic coding exploits dependencies among residual-quantization indices to reduce lossless coding rate. An offline n-gram model supplies symbol probabilities, and arithmetic coding encodes the sequence using a shared model.

B. Context-adaptive Arithmetic Coding

  • The n-gram model exploits redundancy in the residual-quantization index sequence, whose neighboring indices are strongly correlated.Arithmetic coding uses conditional entropy as the rate bound and an estimated model whose cross-entropy determines expected coding rate.Source: B. Context-adaptive Arithmetic Coding · p. 5
  • Conditioning on neighboring indices lets the n-gram estimator better approximate sequence statistics and improve lossless compression efficiency.The paper uses an n-gram model to keep computational overhead low for real-time visual transmission.Source: B. Context-adaptive Arithmetic Coding · p. 5
  • Additive smoothing assigns nonzero probabilities to unseen n-gram combinations, preventing interval collapse during arithmetic coding.The smoothed model and frequency table are built offline and shared by encoder and decoder, requiring no additional side information.Source: B. Context-adaptive Arithmetic Coding · p. 5
  • Arithmetic coding uses the estimated conditional probabilities to progressively narrow an interval until it represents the entire index sequence.The resulting interval provides a compact binary representation for transmission, and the decoder recovers the same index sequence using the shared model.Source: B. Context-adaptive Arithmetic Coding · pp. 5–6

V. ADAPTIVE MODEL SELECTION UNDER CHANNEL CONSTRAINTS

Runtime adaptive model selection chooses among residual-quantization depths using channel state and system constraints.

V. ADAPTIVE MODEL SELECTION UNDER CHANNEL CONSTRAINTS

  • The adaptive depth controller selects semantic encoding depth according to observed channel state and system constraints before transmission.The selection is formalized as a channel-aware model-selection problem over a discrete set of residual-quantization depths.Source: §V · p. 6

A. Scaling Dimensions and Trade-offs

The framework trades reconstruction quality and transmission cost through residual-quantization depth, while tuning entropy-model order offline. At runtime, it selects a depth using profiled quality and expected latency under channel and deadline constraints.

A. Scaling Dimensions and Trade-offs

  • Deeper residual quantization preserves finer semantic details and improves reconstruction quality, but increases latent size and transmission and decoding cost.Shallower depths yield more compact representations and lower latency at reduced semantic fidelity; Fig. 3 illustrates clearer details and vehicle boundaries at greater depth.Source: A. Scaling Dimensions and Trade-offs · p. 6
  • For RQ-NAC(8), moving from 1-gram to 2-gram raises encoding latency from 47.2 ms to 77.9 ms, while 3-gram requires 4.4 s, 210.7 GiB, and 22.8 GB.The authors therefore use a moderate-order entropy model for real-time deployment and tune its order offline.Source: A. Scaling Dimensions and Trade-offs · p. 6, Table I · p. 6
  • The system uses residual-quantization depth for runtime adaptation and entropy-model order as an offline parameter, adjusting bitrate and fidelity without codec retraining.The entropy-model trade-off reflects increasing context coverage alongside exponential context-space growth, data sparsity, and rising storage and lookup overhead.Source: A. Scaling Dimensions and Trade-offs · p. 6

B. Channel-Aware Adaptive Model Selection

  • At each transmission instance, the encoder selects a candidate depth using effective multicast rate and latency budget, maximizing utility subject to the deadline constraint.Semantic quality and expected end-to-end latency are profiled offline for candidate depths and stored for runtime lookup; no online quality estimation or ground-truth reference is required.Source: B. Channel-Aware Adaptive Model… · pp. 6–7
  • If no candidate depth meets the deadline, the system falls back to the depth with minimum expected end-to-end latency.Otherwise, it maximizes utility over the feasible set.Source: B. Channel-Aware Adaptive Model… · p. 7
  • The runtime selection procedure has O(|D|) complexity and requires no retraining or iterative optimization.It selects from a small candidate set before transmission under the stochastic block-level HARQ process.Source: B. Channel-Aware Adaptive Model… · p. 7

C. Safety-Critical Perception with Domain Adaptation

The section motivates downstream detection as a safety-relevant test of reconstructed visual representations and describes domain-adaptive detector fine-tuning for vehicular scenes.

C. Safety-Critical Perception with Domain Adaptation

  • Nearby-vehicle detection matters for collision avoidance and immediate decision-making, so evaluation extends beyond reconstruction fidelity to downstream object detection.The section identifies gaps between generic natural-image data and vehicular scenes, and between original and reconstructed images.Source: C. Safety-Critical Perception with… · p. 7
  • The domain-adaptive detection example initializes YOLOv8n from COCO-pretrained weights and trains with the standard YOLOv8 detection loss.The loss combines box, classification, and distribution focal terms; the box term uses Complete IoU.Source: C. Safety-Critical Perception with… · p. 7
  • The detector is fine-tuned on vehicle-centric original images to capture driving-scene structures and generalize across reconstructed images without compression-specific retraining.The approach does not fine-tune the detector on compressed images.Source: C. Safety-Critical Perception with… · p. 7

VI. EVALUATION METHODOLOGY

The evaluation uses dashcam images and compression baselines to test RQ-NAC, with implementation choices defining its quantization and entropy-coding configurations.

A. Datasets and Baselines

  • The 100K Real-World Vehicle Dashcam Image Dataset supplies 20,000 training images and 1,000 test images, with a separate 1,000-frame Car Crash Dataset case study for distribution-shift evaluation.Unless otherwise specified, RQ-NAC and VQ-VAE use a 5,000-image training subset; RQ-NAC(4) is also trained on all 20,000 images.Source: A. Datasets and Baselines · pp. 7–8
  • RQ-NAC is compared with VQ-VAE, DiffEIC, QRes-VAE, HEVC, and JPEG using 1,280×720 inputs padded to 1,280×736.BPP calculations use the original image dimensions.Source: A. Datasets and Baselines · p. 8

B. Implementation Details

  • The RQ-VAE-based encoder maps RGB images to 256-dimensional latents, and residual-quantization depth sets the number of refinement stages and bitrate–fidelity operating points.The codebook has 2,048 entries, and the decoder progressively upsamples the quantized representation.Source: B. Implementation Details · p. 8
  • The default NAC configuration uses a 2-gram context model with smoothing coefficient α = 0.1.Models were trained on eight NVIDIA A100 GPUs; runtime profiling used an Intel Core Ultra 9 275HX CPU and NVIDIA RTX 5080 Laptop GPU.Source: B. Implementation Details · p. 8

C. Evaluation Metrics

The evaluation defines image-quality, compression, communication, and perception metrics, then simulates multicast conditions with an abstract PHY and sampled scenarios.

C. Evaluation Metrics

  • SSIM, PSNR, LPIPS, and FID characterize reconstructed-image structure, pixel accuracy, perceptual similarity, and feature-space distributional distance, respectively.FID compares the means and covariances of reference and reconstructed feature representations.Source: C. Evaluation Metrics · p. 8
  • BPP measures bits per original pixel and compression ratio compares uncompressed with compressed data sizes; JPEG source files are references, so reconstruction metrics are not reported for JPEG.The section also defines success ratio, deadline-meet ratio, and end-to-end latency, including source encoding, wireless transmission, and destination decoding.Source: C. Evaluation Metrics · p. 8
  • Object detection is assessed with precision, recall, mAP, and WDS, which weights correctly detected objects by normalized bounding-box area.A detection counts as correct when the same-class prediction has IoU ≥0.5.Source: C. Evaluation Metrics · pp. 8–9

D. PHY Abstraction and Simulation Framework

  • The PHY abstraction maps effective SINR and allocated resources to successfully decoded bits per scheduling interval, with link adaptation targeting a specified BLER and bounded HARQ retransmissions.Monte Carlo scenarios sample multicast group size, user SINRs, and concurrent flows; quality and latency weights are 0.8 and 0.2.Source: D. PHY Abstraction and Simulation… · p. 9
  • The channel model is a system-level abstraction, not a time-resolved mobility model, and omits correlated fast fading, Doppler, detailed interference, and CSI estimation or feedback delay.Depth selection assumes a current effective SINR estimate is available.Source: D. PHY Abstraction and Simulation… · p. 9

VII. EXPERIMENTAL RESULTS

The experiments assess RQ-NAC’s reconstruction and compression trade-offs across quantization depths and training scales, then examine context-adaptive entropy coding.

1) Performance of Residual Quantization:

  • At approximately 0.04 BPP, RQ-NAC(4)-20k achieves an SSIM of 0.7779 and an LPIPS of 0.1541, with substantially lower LPIPS than the considered learned baselines and HEVC.At that bitrate it maintains competitive SSIM; HEVC requires considerably higher bitrate to approach a similar perceptual-quality level.Source: Figure 4 · p. 9, 1) Performance of Residual Quantization · p. 10
  • RQ-NAC(16) achieves SSIM 0.7832, PSNR 25.00 dB, and LPIPS 0.1519, outperforming VQ-VAE on SSIM and PSNR while compression ratio falls from 546.40× at depth 4 to 136.60× at depth 16.The comparison is among learned models trained on the 5k subset.Source: 1) Performance of Residual Quantization · p. 9
  • RQ-NAC(4)-20k achieves reconstruction performance comparable to RQ-NAC(16)-5k, indicating that more training data can improve quality with lower bitrate overhead.Increasing depth progressively improves visual fidelity, including sharper edges and clearer object contours.Source: 1) Performance of Residual Quantization · pp. 9–10, Figure 5 · p. 10

2) Performance of Entropy-based Lossless Compression:

  • The 2-gram context model raises entropy-coding compression ratio to around 1.3× from approximately 1.15× with arithmetic coding alone, while 3-gram coding degrades efficiency under limited-data context sparsity.Higher-order contexts also increase frequency-table storage and lookup costs.Source: 2) Performance of Entropy-based… · pp. 10–11
  • With α = 0.1, 2-gram modeling achieves up to 671.09× compression at quantization depth 4, with high perceptual quality.Across tested smoothing values, compression differences are marginal; α = 0.1 slightly outperforms the alternatives at depths 4 and 8.Source: 2) Performance of Entropy-based… · p. 11

B. Adaptive Performance under Dynamic Channel Conditions

The evaluation profiles processing cost and tests whether runtime depth selection responds to scenario and communication conditions, including latency and multicast coverage requirements.

1) Profiling Results and System Efficiency:

  • Semantic encoding remains approximately 120 ms across RQ depths because the convolutional backbone runs once per image, while NAC decoding cost grows with longer latent index sequences.The pipeline remains at a hundreds-of-milliseconds latency level even at higher depths.Source: 1) Profiling Results and System… · p. 11, Table VI · p. 11

2) Performance of Model Adaptation:

  • The controller selects deeper RQ configurations in Cruise, intermediate depth in Junction, and shallow depth in Collision, tracking the changing multicast load and latency constraints.The selection adapts without model retraining or online optimization.Source: 2) Performance of Model Adaptation · p. 11, Table VI · p. 11

3) Sensitivity Analysis:

  • Within Junction, greater latency budgets, better channel quality, and fewer concurrent flows shift selection toward deeper quantization, while receiver-population changes have comparatively smaller effects under the coverage-quantile abstraction.These controlled variations show that selection responds to communication conditions rather than scenario labels.Source: Figure 8 · p. 12, 3) Sensitivity Analysis · p. 12
  • Under the stringent Collision setting, RQ-NAC provides higher timely-delivery reliability and lower end-to-end latency than JPEG; increasing target receiver coverage lowers deadline-meet ratios, but RQ-NAC remains ahead across tested targets.The relative reliability advantage is reported across the tested multicast coverage requirements.Source: 3) Sensitivity Analysis · p. 12, Figure 9 · p. 12

C. Perception Performance with Semantic Applications

Deeper RQ-NAC reconstructions improve downstream vehicle detection and preserve higher-importance objects, while unseen accident scenes retain key structures and the fixed entropy model remains effective. The conclusion reports improved delivery reliability and timely transmission alongside low-latency operation and robust detection.

C. Perception Performance with Semantic Applications

  • After fine-tuning on domain-specific vehicular data, RQ-NAC(16) reaches 0.880 mAP50 and surpasses the original baseline in precision, while original JPEG retains the highest overall mAP50 and mAP50−95.Source: C. Perception Performance with… · p. 13
  • 0.897 WDS: RQ-NAC at depth 16 remains close to the original baseline, and WDS improves monotonically with reconstruction depth.The results indicate better preservation of objects assigned higher importance scores.Source: C. Perception Performance with… · p. 13

D. Robustness under Distribution Shift

VIII. CONCLUSION

  • The system-level evaluation reports improved delivery reliability and timely transmission versus traditional image transmission, with low-latency operation and robust object detection.The conclusion also reports that adaptive semantic compression can mitigate channel variability and network congestion across diverse vehicular application environments.Source: §VIII · p. 13
Loading 2610.10886v1…