---
title: 'SlimComm: Sparse 3D Perception in CAVs'
url: https://www.emergentmind.com/topics/slimcomm
type: topic
---

# SlimComm: Sparse 3D Perception in CAVs

SlimComm is a communication-efficient framework for cooperative 3-D perception in connected autonomous vehicles (CAVs) that integrates 4-D radar Doppler with a query-driven sparse communication scheme. Its central objective is to avoid transmitting dense Bird’s-Eye-View (BEV) feature maps, which can overwhelm inter-vehicle bandwidth, while preserving detection accuracy under occlusion and limited sensor range. The framework constructs a motion-centric dynamic map, generates two classes of sparse BEV queries, exchanges only query-specific features, and fuses collaborator responses through multi-scale gated deformable attention. The same work also introduces the CARLA-based OPV2V-R and Adver-City-R datasets with per-point Doppler radar for evaluation [2508.13007].

## 1. System formulation and architectural organization

Each CAV \(i\) collects a time-synchronized LiDAR point cloud \(L_i\) and six 4-D radar point clouds \(R_i\), where each radar point carries \((x,y,z)\) plus Doppler velocity. LiDAR and radar points are pillarized into two BEV tensors, \(\bm L_i\) and \(\bm R_i\), concatenated, and passed through a ResNet-style BEV backbone. For each scale \(l\), the encoder produces a multi-scale feature map \(\bm F_{i,l}\in\mathbb R^{C_l\times H_l\times W_l}\), together with three semantic priors on the highest-resolution grid \((H\times W)\): a Dynamic Map \(\bm D_i\in\{0,1\}^{H\times W}\), a Confidence Map \(\bm C_i\in[0,1]^{H\times W}\), and a Foreground Density Map \(\bm V_i\in[0,1]^{H\times W}\) [2508.13007].

The overall processing chain is explicitly modular. An Ego Query Generator uses \(\bm D_{i,l}\), \(\bm C_{i,l}\), and \(\bm V_{i,l}\) to emit sparse BEV queries at each scale. The communicator then broadcasts the sparse query locations \(\bm Q_{i,l}\) and the ego pose. Each neighboring vehicle \(j\) first warps its foreground-density prior into the ego frame and decides whether to participate; if any ego query falls on a cell satisfying \(\bm V_{j\to i}(u,v)>\tau_{\mathrm{com}}\), the neighbor warps its multi-scale features into ego coordinates, extracts a \(3\times 3\) halo around each queried location, and transmits only those halo-enriched features \(\bm H_{j\to i,l}\). Gated Multi-Scale Deformable Fusion merges ego features and collaborator responses into fused features \(\tilde{\bm F}_{i,l}\), which are up-sampled and aggregated into a unified BEV tensor \(\tilde{\bm S}_i\) for a lightweight 3D detection head. This organization makes sparse communication a first-class design principle rather than a post hoc compression stage.

## 2. Doppler-guided motion-centric dynamic map

A defining component of SlimComm is its use of radar Doppler to separate moving from static content before query generation. Let \(\mathbf v^{\mathrm{veh}_i}\in\mathbb R^3\) denote ego velocity in ego coordinates. For radar \(k\) with extrinsic rotation \(\mathbf R_{i,k}\), the ego velocity in radar-\(k\) coordinates is

\[
\mathbf v^{\mathrm{veh}_{i,k}} = \mathbf R_{i,k}\,\mathbf v^{\mathrm{veh}_i}\,.
\]

Each radar returns an absolute Doppler measurement \(v^{\mathrm{Doppler}}_{i,k,n}\) along the line of sight \(\mathbf u_{i,k,n}=\mathbf p_{i,k,n}/\|\mathbf p_{i,k,n}\|\). After compensating for own-vehicle motion, the radial velocity of point \(n\) is

\[
v^{\mathrm{radial}}_{i,k,n}
= v^{\mathrm{Doppler}}_{i,k,n}
+ \bigl(\mathbf v^{\mathrm{veh}_{i,k}}\cdot\mathbf u_{i,k,n}\bigr)\,.
\]

The fused LiDAR-radar point cloud is then discretized into a BEV grid of size \(H\times W\). A grid cell \((u,v)\) is marked dynamic if there exists a radar point projected into that cell whose compensated radial speed exceeds the threshold \(v_{\mathrm{th}}=1.0\,\mathrm{m/s}\):

\[
\exists\;k,n \quad\text{s.t.}\quad
(x_{i,k,n},y_{i,k,n})\in\text{cell}(u,v)
\quad\text{and}\quad
\bigl|v^{\mathrm{radial}}_{i,k,n}\bigr| > v_{\mathrm{th}}\,.
\]

The resulting binary map is

\[
\bm D_i(u,v)=
\begin{cases}
1, & \text{cell \((u,v)\) dynamic},\\
0, & \text{otherwise.}
\end{cases}
\]

This motion-centric construction gives the communication module a physically grounded prior for prioritizing moving objects. Because the six radars provide full \(360^\circ\) coverage and each radar point includes per-point Doppler, the method is designed to support moving/static separation across all directions [2508.13007].

## 3. Query taxonomy and two-stage offset mechanism

SlimComm generates two query types at each BEV scale \(l\), each containing \(N_l^r\) reference points. The first type is **Heuristic Reference Points (HRP)**. Here, \(\bm D_i\) and \(\bm C_i\) are down-sampled to \(\bm D_{i,l}\) and \(\bm C_{i,l}\). Pool 1 contains all cells with \(\bm D_{i,l}(u,v)=1\); Pool 2 contains the highest-confidence cells not already in Pool 1 until \(|\text{Pool 1}\cup\text{Pool 2}|=N_l^r\). The selected coordinates are stored in \(\bm R_{i,l}^h\in\mathbb R^{N_l^r\times 2}\), and the associated embeddings are obtained by bilinear sampling:

\[
\bm E_{i,l}^h=\mathrm{BilinearSample}(\bm F_{i,l},\bm R_{i,l}^h)\in\mathbb R^{N_l^r\times C_l}.
\]

The second type is **Exploratory Reference Points (ERP)**, intended to probe occluded regions. Occluder peaks are first found in \(\bm C_{i,l}\) by

\[
\bm C_{i,l}^o
=
\bigl(\mathrm{MaxPool}_{3\times 3}(\bm C_{i,l})\bigr)\;\wedge\;\bm C_{i,l},
\]

followed by thresholding with the per-scene percentile \(p_l\). For each peak, a stochastic shadow offset \(\Delta^{\mathrm{shadow}}\) is sampled, yielding \(\bm R_{i,l}^e\in\mathbb R^{N_l^r\times 2}\). Each ERP is embedded as

\[
\bigl[\,
\bm F_{i,l}(u,v)\;\|\;\Delta^{\mathrm{shadow}}\;\|\;\bm t_l
\bigr]\in\mathbb R^{2C_l+2},
\]

and then mapped by a two-layer MLP to \(\bm E_{i,l}^e\in\mathbb R^{N_l^r\times C_l}\) [2508.13007].

The two query branches are combined through anchor formation and offset prediction. The anchor set is

\[
A_{i,l} = \bm R_{i,l}^h\;\cup\;\bm R_{i,l}^e,
\qquad |A_{i,l}|=2N_l^r.
\]

A coarse MLP predicts a 2-D offset \(O_{i,l}\in\mathbb R^{2N_l^r\times 2}\), producing nudged centers

\[
\widetilde A_{i,l} = A_{i,l} + O_{i,l}.
\]

To prevent HRP and ERP from collapsing to similar locations, SlimComm introduces an auxiliary margin loss:

\[
\mathcal L_{\mathrm{offset}}
= \sum_l
\Bigl[
\delta_l
-
\bigl(
\mathbb E\|O_{i,l}^{\mathrm E}\|_2
-
\mathbb E\|O_{i,l}^{\mathrm H}\|_2
\bigr)
\Bigr]_+,
\]

where \(\delta_l\) is the mean occluder-to-shadow distance at scale \(l\). A deformable-attention head then predicts 9 local offsets \(\{\Delta^{\mathrm{fine}}_{h,p}\}\) for each nudged anchor, producing final query locations

\[
\bm Q_{i,l}
=
\bigl\{
\tilde a + \Delta^{\mathrm{fine}}_{h,p}
\bigr\}_{\substack{\tilde a\in\widetilde A_{i,l}\\1\le h\le N_l^r,\;1\le p\le 9}}.
\]

A common misconception in sparse cooperative perception is that sparsification is synonymous with selecting only already-visible or already-confident locations. SlimComm explicitly departs from that assumption by coupling dynamic/high-confidence HRP with occlusion-oriented ERP. This suggests that its sparsity pattern is designed not only for compression but also for targeted exploration of hidden structure.

## 4. Sparse communication protocol and gated deformable fusion

The communication stage is query-centric. The ego vehicle broadcasts \(\{\bm Q_{i,l}, \text{pose}_i\}\). Each neighbor \(j\) warps its foreground density map into the ego frame, forming \(V_{j\to i}\), and opts in only if any query at the highest-resolution scale lands on a location with \(V_{j\to i}(u,v)>\tau_{\mathrm{com}}\). For participating neighbors, the warped multi-scale features \(F_{j,l}^{\mathrm{warp}}\) are used to extract halo-enriched responses \(H_{j\to i,l}\) by taking a \(3\times 3\) neighborhood around each ego query. Only these query-conditioned local features are transmitted [2508.13007].

Fusion proceeds in three steps. First, collaborator features are averaged:

\[
\bm H_{i,l}^{\mathrm{agg}}
=
\frac{1}{N}\sum_{j=1}^{N}\bm H_{j\to i,l},
\]

where \(\bm H_{j\to i,l}\in\mathbb R^{N_l^q\times 9C_l}\). Second, deformable cross-attention is applied using the ego query embeddings \(\bm E_{i,l}=[\bm E_{i,l}^h;\,\bm E_{i,l}^e]\), the final query locations \(\bm Q_{i,l}\), and the aggregated collaborator halos:

\[
\bm F_{i,l}^{\mathrm{CAV}}
=
\mathrm{Scatter}\Bigl(
\mathrm{DeformAttn}\bigl(
\bm E_{i,l},\,\bm Q_{i,l},\,\bm H_{i,l}^{\mathrm{agg}}
\bigr),
\bm Q_{i,l}
\Bigr).
\]

All unqueried grid cells are set to zero. Third, the collaborator-informed features are blended with the ego features by gated residual fusion:

\[
\tilde{\bm F}_{i,l}
=
\bigl(1-\bm G_{i,l}\bigr)\odot \bm F_{i,l}
+
\bm G_{i,l}\odot \bm F_{i,l}^{\mathrm{CAV}},
\]

with gate

\[
\bm G_{i,l}
=
\sigma\Bigl(
\mathrm{Conv}_{1\times 1}\bigl[
\bm F_{i,l}\,\|\,\bm F_{i,l}^{\mathrm{CAV}}
\bigr]
\Bigr)\in[0,1]^{C_l\times H_l\times W_l}.
\]

The fused feature pyramid \(\{\tilde{\bm F}_{i,l}\}\) is then up-sampled and aggregated into \(\tilde{\bm S}_i\) for detection. The end-to-end pseudocode supplied with the method makes clear that query generation, neighbor selection, halo extraction, deformable attention, and gated residual blending are integrated in a single loop over time steps and scales.

## 5. Datasets and evaluation setting

SlimComm is evaluated on two CARLA-based cooperative perception datasets, OPV2V-R and Adver-City-R, both built on CARLA and OpenCDA, recorded at 10 Hz, and annotated with 3-D bounding boxes. Each CAV uses the same sensor suite: one 64-channel LiDAR with \(1.3\) M points/s and \(120\) m range, six 4-D radars with \(0.06\) M points/s each and \(150\) m range, four RGB cameras providing \(360^\circ\) FOV, and GPS/IMU. The radar returns \((x,y,z,v_{\mathrm{Doppler}})\), and the six-radar layout provides full \(360^\circ\) coverage [2508.13007].

| Dataset | Scene characterization | Traffic statistics |
|---|---|---|
| OPV2V-R | “General urban” | Avg. neighbors per frame: 2 (up to 5); avg. vehicles in range \(\approx 16.7 \pm 7.9\) |
| Adver-City-R | “Crash-relevant scenarios” | Avg. neighbors per frame: 2; avg. vehicles \(\approx 20.8 \pm 7.3\); five road types; scenario lengths 23–39 s; densities 10–33 vehicles |

The two datasets serve distinct roles. OPV2V-R represents general urban operation, whereas Adver-City-R emphasizes crash-relevant scenarios and explicitly stratifies results by General, Sparse, and Dense traffic settings. Because both datasets include per-point Doppler radar, they support direct evaluation of the motion-centric dynamic map rather than requiring radar to be treated only as an auxiliary geometric sensor.

## 6. Quantitative performance and research significance

Bandwidth is computed in float32 as

\[
B_s = 4\sum_l N_{s,l}\quad[\mathrm{bytes}],
\qquad
\overline{\mathrm{BD}}=
\frac{1}{|\mathcal S|}
\sum_{s\in\mathcal S}
\frac{B_s}{10^6}
\quad[\mathrm{MB/frame}].
\]

Relative to full-map sharing of a dense BEV tensor, SlimComm uses only \(0.5\)–\(1.1\) MB/frame, corresponding to up to \(90\%\) lower bandwidth. On Adver-City-R, SlimComm reports AP@0.5 of \(0.67/0.72/0.65\) and AP@0.7 of \(0.54/0.63/0.52\) for General/Sparse/Dense, with BD \(\approx 1.13\) MB/frame. In the same setting, Where2Comm reports \(0.47/0.47/0.47\) and \(0.23/0.33/0.29\) with BD \(\approx 6.20\) MB, and AttFusion reports \(0.64/0.69/0.63\) and \(0.47/0.54/0.46\) with BD \(\approx 13.47\) MB. On OPV2V-R, SlimComm reports AP@0.5 \(=0.87\), AP@0.7 \(=0.80\), and BD \(\approx 0.63\) MB/frame, compared with Where2Comm at \(0.86/0.77\) and BD \(\approx 4.45\) MB, and AttFusion at \(0.89/0.80\) and BD \(\approx 6.72\) MB [2508.13007].

These results are summarized below.

| Setting | SlimComm | Baselines |
|---|---|---|
| Adver-City-R | AP@0.5 \(=0.67/0.72/0.65\); AP@0.7 \(=0.54/0.63/0.52\); BD \(\approx 1.13\) MB/frame | Where2Comm: BD \(\approx 6.20\) MB; AttFusion: BD \(\approx 13.47\) MB |
| OPV2V-R | AP@0.5 \(=0.87\); AP@0.7 \(=0.80\); BD \(\approx 0.63\) MB/frame | Where2Comm: BD \(\approx 4.45\) MB; AttFusion: BD \(\approx 6.72\) MB |

Within the reported experiments, SlimComm matches or exceeds the best baselines under all densities while using only \(\approx 10\%\) of their bandwidth. A plausible implication is that the combination of Doppler-guided motion priors, exploratory occlusion probing, and query-conditioned halo transmission changes the usual trade-off between communication volume and cooperative detection accuracy. Rather than treating bandwidth reduction as a purely compression-oriented objective, SlimComm frames it as a structured feature-selection problem over dynamic and occluded BEV regions.

Source: https://www.emergentmind.com/topics/slimcomm