ReVQom: Efficient Codec for Collaborative Perception

• ReVQom is a learned feature codec for multi-agent collaborative perception that compresses spatially-detailed intermediate features using discrete index transmission.
• It employs a spatial bottleneck network and multi-stage residual vector quantization to achieve up to 1365x compression while retaining essential detection accuracy.
• The method enables ultra-low-bandwidth communication for connected agents like autonomous vehicles and robots, ensuring effective BEV-based fusion and detection.

Searching arXiv for the cited ReVQom paper and closely related collaborative-perception baselines to ground the article. ReVQom is a learned feature codec for multi-agent collaborative perception (CP) that preserves spatial identity while compressing intermediate features for transmission across connected agents such as autonomous vehicles, unmanned aerial vehicles, and robots. It is presented as an end-to-end method that compresses feature dimensions via a simple bottleneck network followed by multi-stage residual vector quantization (RVQ), so that only per-pixel code indices are transmitted rather than full intermediate features (Shenkut et al., 25 Sep 2025). Within BEV-based CP, the method is designed to reduce the communication burden of sharing rich intermediate representations while retaining the spatial structure needed for downstream fusion and detection. On the DAIR-V2X real-world CP dataset, ReVQom reports compression from 8192 bits per pixel (bpp) of uncompressed 32-bit float features to 6–30 bpp per agent, corresponding to 273x compression at 30 bpp and 1365x compression at 6 bpp, with minimal accuracy loss (Shenkut et al., 25 Sep 2025).

1. Problem setting and design objective

Multi-agent collaborative perception seeks to improve scene understanding by sharing information across connected agents. In the stated application setting, collaboration improves detection especially in occluded or long-range scenarios by fusing information from different vantage points, including roadside units and vehicles (Shenkut et al., 25 Sep 2025). The central systems problem is that communication bandwidth constrains scalability: transmitting high-dimensional Bird’s Eye View features from each agent can exceed practical V2X channel budgets.

ReVQom is proposed to address this bottleneck by compressing spatial feature representations aggressively enough to enable ultra-low-bandwidth communication without sacrificing collaborative perception quality (Shenkut et al., 25 Sep 2025). The method specifically targets the case where raw feature transmission is prohibitively expensive: the paper identifies uncompressed features as 8192 bpp, arising from 32 bits times 256 channels, and contrasts this with compressed operating points between 6 and 30 bpp (Shenkut et al., 25 Sep 2025). This framing places ReVQom within the broader line of BEV-based collaborative perception systems such as CoBEVT, F-Cooper, V2X-ViT, and Where2comm, which are discussed in the reported comparisons (Shenkut et al., 25 Sep 2025).

A notable aspect of the formulation is that ReVQom is not described as a generic scene codec. It is a feature codec tailored to intermediate representations used for multi-agent fusion. The emphasis on preserving spatial identity distinguishes the method from approaches that trade away the explicit spatial arrangement of features, which the paper argues is crucial for downstream fusion and detection in BEV (Shenkut et al., 25 Sep 2025).

2. Architectural formulation

At the architectural level, ReVQom consists of two principal components: a spatial bottleneck network and a multi-stage RVQ module (Shenkut et al., 25 Sep 2025). The bottleneck first reduces channel dimensionality while preserving the spatial layout. Concretely, each agent applies a $1\times1$ convolution to map the BEV feature tensor from $C$ channels to $C_r$, with $C_r \ll C$, followed by Group Normalization (Shenkut et al., 25 Sep 2025). Because the operation is pointwise in the spatial dimensions, the $H \times W$ layout is preserved.

The reduced feature at each pixel location is then processed by multi-stage residual vector quantization. Each stage uses a learned codebook of size $K$. The quantizer searches for the nearest codeword to the current residual, beginning with the bottleneck feature itself, and encodes its index. The selected codeword is subtracted from the residual, and the process repeats for the prescribed number of stages $n_q$ (Shenkut et al., 25 Sep 2025). This residual decomposition allows the transmitted representation to be a sequence of discrete indices per pixel rather than a dense floating-point vector.

The sender-receiver protocol is correspondingly simple. The sender takes an input BEV feature map $F$ of size $H \times W \times C$, applies $1\times1$ Conv + GroupNorm to obtain $C$0 of size $C$1, then for each pixel and RVQ stage finds the nearest codebook entry and transmits the codebook index $C$2 (Shenkut et al., 25 Sep 2025). The output is therefore a per-pixel sequence of indices $C$3.

At the receiver, reconstruction is obtained by summing the corresponding stagewise codewords,

$C$4

then applying a post-affine transformation, described as another $C$5 convolution followed by ReLU, and finally expanding channels back to the original size $C$6 to recover a reconstructed BEV feature for fusion (Shenkut et al., 25 Sep 2025). Codebooks are updated during training using Exponential Moving Average (EMA), which the paper describes as providing stability and consistency across agents (Shenkut et al., 25 Sep 2025).

3. Rate, compression, and transmitted representation

The communication model is expressed directly in terms of codebook indices. Because each pixel transmits one index per RVQ stage, the compressed bit rate in bits per pixel is

$C$7

and the transmission cost per agent is

$C$8

The uncompressed reference cost is $C$9, reflecting 32-bit floating-point features with $C_r$0 channels (Shenkut et al., 25 Sep 2025).

The compression ratio is correspondingly defined as

$C_r$1

with uncompressed bpp given by $C_r$2 and compressed bpp given by $C_r$3 (Shenkut et al., 25 Sep 2025). For the reported BEV setting, the uncompressed baseline is 8192 bpp, and the compressed operating points are as follows (Shenkut et al., 25 Sep 2025):

Setting	bpp	Compression
ReVQom-$C_r$4 ($C_r$5)	6	1365x
ReVQom-T ($C_r$6)	12	683x
ReVQom-S ($C_r$7)	18	455x
ReVQom-M ($C_r$8)	24	341x
ReVQom-L ($C_r$9)	30	273x

These figures encapsulate the core operating trade-off of the method: the communication payload is controlled by the number of RVQ stages and codebook size, while detection accuracy depends on whether the resulting reconstruction preserves the downstream semantics needed for collaborative fusion (Shenkut et al., 25 Sep 2025). A plausible implication is that ReVQom treats communication as a structured discrete coding problem rather than as sparse raw-feature selection.

4. Experimental evaluation on DAIR-V2X

The empirical study is conducted on the DAIR-V2X dataset, described as a multi-view LiDAR benchmark with vehicle and roadside-unit observations and ground-truth 3D labels (Shenkut et al., 25 Sep 2025). ReVQom is integrated into CoBEVT as a modular compressor (Shenkut et al., 25 Sep 2025). The principal detection metrics reported are $C_r \ll C$0 and $C_r \ll C$1.

The paper compares ReVQom against prior CP methods with substantially higher bandwidth. The reported results are:

Method	bpp	Compression	[email protected]	[email protected]
F-Cooper (raw)	8192	1x	0.704	0.648
V2X-ViT	6144	1.3x	0.745	0.676
Where2comm	512	16x	0.701	0.634
ReVQom-$C_r \ll C$2	6	1365x	0.690	0.558
ReVQom-T	12	683x	0.699	0.609
ReVQom-S	18	455x	0.747	0.651
ReVQom-M	24	341x	0.753	0.666
ReVQom-L	30	273x	0.725	0.636

Several points are explicit in these results. First, at 18 bpp, ReVQom-S achieves 455x compression with $C_r \ll C$3 $C_r \ll C$4, which the paper states outstrips all prior methods, including methods using essentially uncompressed features (Shenkut et al., 25 Sep 2025). Second, at 24 bpp, ReVQom-M attains the highest reported $C_r \ll C$5 of $C_r \ll C$6 and $C_r \ll C$7 of $C_r \ll C$8 among the listed methods (Shenkut et al., 25 Sep 2025). Third, even the ultra-low-bandwidth settings of 6–12 bpp remain competitive, with $C_r \ll C$9 between $H \times W$0 and $H \times W$1, which the paper characterizes as graceful degradation (Shenkut et al., 25 Sep 2025).

The article’s central comparative claim is therefore not merely that ReVQom compresses more strongly, but that at moderate compressed rates it matches or exceeds the collaborative detection accuracy of earlier systems while operating at one to two orders of magnitude lower transmission budgets (Shenkut et al., 25 Sep 2025).

5. Codebook behavior and ablation findings

The ablation study isolates several design choices: EMA decay rate, number of quantization stages, channel reduction ratio, and codebook size (Shenkut et al., 25 Sep 2025). These analyses are presented as technical evidence for the selected operating configuration rather than as independent algorithmic variants.

For EMA, the paper reports that a lower decay rate of $H \times W$2 outperforms the standard $H \times W$3, which is interpreted as indicating that faster codebook adaptation helps in highly dynamic, multi-agent settings (Shenkut et al., 25 Sep 2025). For quantization depth, three-stage RVQ, $H \times W$4, gives top performance, while more stages, such as four, produce diminishing returns (Shenkut et al., 25 Sep 2025). For channel reduction, the best result is obtained at reduction ratio $H \times W$5, while more aggressive compression degrades accuracy sharply (Shenkut et al., 25 Sep 2025). For codebook size, $H \times W$6 is identified as optimal in balancing bandwidth and accuracy, whereas $H \times W$7 is reported to lead to overfitting and reduced AP (Shenkut et al., 25 Sep 2025).

The paper also analyzes codebook assignment patterns. At low $H \times W$8, most indices encode only background; for example, code 0 covers 96–98% of space, while the remaining codes are used for foreground semantics (Shenkut et al., 25 Sep 2025). This observation is presented as evidence that the codec leverages both spatial and channel redundancy. In practical terms, large portions of the BEV map can be represented by a small set of background assignments, while semantically salient regions consume the higher-information code usage. This suggests that the gains are not solely attributable to numeric compression, but to the alignment between scene structure and discrete code allocation.

A common misconception would be to interpret the highest codebook size as necessarily best. The reported ablations contradict this: $H \times W$9 at 30 bpp underperforms the 24 bpp setting with $K$0, and the paper explicitly attributes this to diminished returns and possible overfitting (Shenkut et al., 25 Sep 2025).

6. Relation to collaborative perception practice and deployment constraints

ReVQom is described as a drop-in replacement for existing BEV-based collaborative networks such as CoBEVT (Shenkut et al., 25 Sep 2025). Its practical appeal derives from the communication protocol: only index streams are transmitted, while codebooks are pre-shared and synchronized across agents (Shenkut et al., 25 Sep 2025). This protocol is well matched to V2X systems where communication bandwidth is limited but maintaining shared model state is feasible.

The deployment discussion in the paper identifies several practical implications. First, the index-based communication scheme allows operation within realistic V2X channel budgets, even as the number of agents scales up (Shenkut et al., 25 Sep 2025). Second, the reported accuracy retention under strong compression makes the method relevant for safety-critical, real-time collaborative applications (Shenkut et al., 25 Sep 2025). Third, orders-of-magnitude bandwidth savings may permit inclusion of more agents and larger operational areas without exceeding communication constraints (Shenkut et al., 25 Sep 2025).

The same discussion also makes clear that these advantages depend on assumptions that can become constraints. Codebook synchronization is required across all agents, and the paper notes that codebook drift or misalignment in deployment requires handling (Shenkut et al., 25 Sep 2025). Likewise, sudden changes in agent population or scene layout may pose difficulties, even though the improved EMA setting is reported to help with dynamic environments (Shenkut et al., 25 Sep 2025). Extreme channel reduction is also explicitly identified as harmful, implying that spatial quantization and channel reduction must be balanced rather than optimized independently (Shenkut et al., 25 Sep 2025).

These points situate ReVQom as a systems-oriented compression method rather than a purely representational innovation. Its relevance depends on the end-to-end communication-fusion loop in collaborative perception, particularly in V2X deployment scenarios.

7. Limitations, interpretation, and research directions

The limitations reported for ReVQom are specific and operational. The first is the requirement for pre-shared and synchronized codebooks across all agents (Shenkut et al., 25 Sep 2025). The second is that extreme reduction in channels severely degrades performance, showing that the bottleneck cannot be made arbitrarily narrow without impairing detection (Shenkut et al., 25 Sep 2025). The third is that dynamic topologies and sudden scene changes could create difficulties despite the benefits of faster EMA adaptation (Shenkut et al., 25 Sep 2025).

The future directions named in the paper follow directly from these constraints. They include dynamic quantization or adaptive codebooks, on-the-fly codebook updates, adaptive bit allocation depending on scene content or agent topology, investigation of transmission errors, packet loss, and latency in the coded index stream, and extension to broader datasets or heterogeneous sensor settings (Shenkut et al., 25 Sep 2025). These are not presented as solved problems.

From the reported evidence, a plausible implication is that ReVQom’s main contribution is to make discrete, index-based feature transmission viable at rates far below those typically used in collaborative perception, while retaining the explicit BEV spatial structure required by downstream fusion. Another plausible implication is that the method’s strongest empirical regime is not the most extreme compression point, but the mid-range operating points, especially 18–24 bpp, where the bandwidth reduction remains several hundredfold and the detection performance is strongest (Shenkut et al., 25 Sep 2025). In that sense, ReVQom is best understood as a practical communication-accuracy trade-off mechanism for multi-agent perception rather than as a universally optimal low-rate codec.