COOPERTRIM: Adaptive Data Selection for Uncertainty-Aware Cooperative Perception

AI summary

Overview

Research area: Cooperative perception for autonomous driving — specifically bandwidth-efficient feature sharing between connected vehicles, spanning semantic segmentation and 3D object detection.

Technical level: Advanced. The paper assumes familiarity with intermediate feature fusion, conformal prediction, quantile gating, cross-attention, and Lagrangian-constrained optimization.

Scope: This paper proposes CooperTrim, a selection framework that uses temporal uncertainty to decide which encoded features a connected vehicle should transmit to its peers, adapting the transmission volume per frame to match scene complexity.

Paper details as reported: arXiv:2602.13287v3 [cs.CV], published 2026-02-07, category Computer Vision. Authors are affiliated with the University of California, Riverside and the New Jersey Institute of Technology.

What This Paper Is About

Cooperative perception lets autonomous vehicles share encoded sensor representations over wireless links so each agent can perceive regions it cannot see directly, but the volume of sensor data far exceeds the available wireless bandwidth. Existing methods reduce traffic with compression or fixed-threshold feature selection, yet the bandwidth demand still strains current wireless technology. CooperTrim attacks the problem differently: instead of transmitting a smaller but static amount of data every frame, it uses temporal context to identify what has changed and shares a variable quantity on demand, based on how uncertain the recipient agent is about its environment.

Key Contributions

1. A proactive, adaptive selection framework. The paper proposes a learning framework that dynamically determines both feature relevance and sharing quantity, described as a departure from static selection frameworks.

2. A conformal temporal uncertainty metric. Relevance is assessed by quantifying temporal uncertainty across frames using a conformal-prediction-inspired quantile gating mechanism with a learnable quantile threshold, combined with an attention mechanism using an adaptive mask threshold for quantity estimation.

3. First selective cooperative semantic segmentation. The authors state this is the first work demonstrating feature selection on cooperative semantic segmentation, a task that demands pixel-level granularity and therefore transmits much larger volumes than object-level detection.

4. An ε-Greedy-inspired training method. Training borrows the ε-Greedy exploration strategy from reinforcement learning, requesting the entire feature set with probability ε and exploiting learned thresholds with probability (1−ε), with a proposition and proof showing the gradient estimator's bias is scaled down by a factor of (1−ε).

Main Findings

• Bandwidth reduction with comparable accuracy: Across multiple open-source cooperative models, CooperTrim achieves up to 80.28% bandwidth reduction for segmentation and 72.52% for 3D detection while maintaining comparable accuracy.

• Improvement over other selection strategies: Relative to other selection strategies, CooperTrim improves IoU by as much as 45.54% with up to 72% less bandwidth.

• Compatibility with compression: Combined with compression strategies, CooperTrim can reduce bandwidth usage to as low as 1.46% without compromising IoU performance — at 32x compression it uses 1.46% versus 3.76% for CoBEVT and 10.62% for AttFuse in the 32x Dynamic setting.

• Applied to three baselines: CooperTrim-CoBEVT, CooperTrim-Attfuse, and CooperTrim-DiscoNet use average bandwidths of 27.9%, 21.07%, and 10.18% respectively (11.16 Mbps, 8.4 Mbps, and 4.07 Mbps) based on a 128x32x32 latent representation, versus 40 Mbps for the originals.

• 3D detection result: On OPV2V and V2V4Real at IoU 0.5 and 0.7, CooperTrim maintains comparable detection accuracy at 27.48% bandwidth usage (72.52% reduction).

• Head-to-head selection comparison (Table 1): CooperTrim reaches 54.03 dynamic, 24.45 static lane, and 44.38 static road IoU at 11.16 Mbps, versus Where2Comm at 8.62 / 20.40 / 36.46 IoU at 39.6 Mbps and SwissCheese at 35.71 / 12.81 / 32.07 IoU at 10 Mbps.

• Training ablation (Table 2): The full EG+CP+FT recipe gives the best dynamic IoU (54.03), while CP+FT alone drops static road accuracy by 8% relative to CooperTrim, Curriculum+CP+FT consumes much more bandwidth (49.25% dynamic, 51.63% static), and EG+SD+FT uses 35.64% bandwidth on dynamic objects versus CooperTrim's 32.04% while scoring 6% lower on static road segmentation.

• Environment adaptation: Bandwidth varies frame to frame with scene complexity. In critical frame ranges (1000–1800 dynamic, 1000–1500 lane, 1200–1400 road) where baseline CoBEVT consistently underperforms, CooperTrim achieves much higher IoU.

• Robustness to localization error: Tested at 0 cm, ±20 cm, and ±1 m during inference; CooperTrim remains robust up to ±20 cm and maintains comparable IoU at ±1 m with stable bandwidth.

• Robustness to communication latency: Tested from 0 ms to 200 ms; no loss up to 50 ms and only marginal IoU drop at 100 ms and 200 ms, with network usage staying consistent.

• Lowest network overhead: Compared against 10 existing baselines spanning compression, feature selection, and agent selection, CooperTrim consumes the lowest bandwidth at 11.6 Mbps.

• Modest system cost: FPS falls from 10.59 (CoBEVT) to 8.58 (CooperTrim) in both static and dynamic configurations, roughly a 2 FPS decrease, attributed to the added uncertainty and cross-attention modules.

Methodology in Plain English

CooperTrim sits on top of existing cooperative perception models rather than replacing them. The ego vehicle encodes its sensor data into a feature map. Instead of judging each frame in isolation, it compares the current frame's features against the fused features from previous frames using an L1 distance — this is the "conformity score," capturing how much the scene understanding has changed.

The paper borrows the idea of conformal prediction but adapts it in two ways: it learns online frame by frame instead of using a fixed calibration dataset, and it estimates uncertainty over components of the regression values rather than intervals over entire regression outputs. A learnable quantile threshold decides which features are relevant enough to consider, and a cross-attention module — where projections of the temporal uncertainty act as queries and feature projections act as keys and values — produces a relevance score per feature.

A second learnable threshold then determines how many of those relevant features to actually request. High scene complexity produces more temporally diverse encodings, higher conformity scores, and therefore more features above threshold; temporal consistency and perception stability produce fewer. The ego broadcasts a request vector for the selected features, responders apply spatial transformation to match the ego's perspective, and the ego blends the received features into its own map before fusion and the task head.

Training is formulated as a constrained optimization problem: minimize perception loss subject to the selected channels matching a bandwidth budget, with the constraint expressed as a percentage of channels corresponding to 1.6 Mbps. This is solved with a Lagrangian formulation where the multiplier λ is adjusted over training — starting at zero, then scaling exponentially every 10th epoch and linearly in between. The ε-Greedy scheme requests all features with probability ε and exploits the learned thresholds otherwise, which the paper argues reduces gradient noise and stabilizes convergence.

Why This Matters

Impact on research. The paper reframes cooperative perception bandwidth reduction as a temporal-uncertainty problem rather than a spatial-confidence or compression problem. Its claim of being the first work on selective cooperative semantic segmentation opens a harder benchmark than detection, since segmentation needs pixel-level granularity. The conformal-inspired online uncertainty metric and the ε-Greedy training formulation are transferable ideas for other bandwidth-constrained multi-agent systems.

Real-world applications.

• Vehicle-to-everything (V2X) deployments where connected vehicles must share perception over congested or unreliable wireless links.
• Autonomous fleet operations in dense urban environments where scene complexity fluctuates sharply and bandwidth must be rationed dynamically.
• Infrastructure-assisted intersections where roadside units broadcast perception to many vehicles simultaneously.
• Cooperative mapping and navigation pipelines that rely on timely exchange of environment representations.

Industry relevance. Practical cooperative perception has been throttled by the mismatch between sensor richness and wireless capacity. CooperTrim's ability to stay robust under ±20 cm localization error and up to 50 ms latency, and to combine with existing compression to reach 1.46% bandwidth usage, targets the operational conditions that real deployments face. The paper reports a code and pretrained model release at https://cisl.ucr.edu/CooperTrim. The work was sponsored by OUSD (R&E)/RT&L under Cooperative Agreement Number W911NF-20-2-0267.

Future Directions

1. Extending beyond the evaluated tasks. The paper instantiates CooperTrim on semantic segmentation and 3D detection; whether the same temporal-uncertainty selection transfers to prediction, mapping, and navigation — all cited as beneficiaries of cooperative perception — is left open.

2. Reducing the computational cost of selection. CooperTrim loses about 2 FPS relative to CoBEVT (10.59 to 8.58) because of the added uncertainty and cross-attention modules. Making the selection overhead cheaper would strengthen the real-time case.

3. Scaling the latency and localization analysis. Robustness is demonstrated up to 50 ms of latency and ±20 cm of localization error, with only marginal degradation at 100–200 ms and ±1 m. Behavior under more severe channel degradation or adversarial communication conditions is not reported.

4. Broader interoperability with compression and other selection methods. CooperTrim is evaluated alongside lossy quantization and lossless reduction at 1x, 8x, and 32x rates, and against 10 baselines, but the space of hybrid selection-plus-compression combinations is not exhausted.

Target Audience

This paper suits researchers and graduate students in cooperative perception, connected and autonomous vehicle systems, and edge/multi-agent communication. It is also relevant to practitioners building V2X deployments who need to understand the bandwidth–accuracy tradeoff, and to readers interested in applying conformal prediction to online or streaming settings rather than fixed calibration datasets. Some background in intermediate feature fusion and attention mechanisms will help, given the technical level of the methodology.

Authors’ abstract

Cooperative perception enables autonomous agents to share encoded representations over wireless communication to enhance each other's live situational awareness. However, the tension between the limited communication bandwidth and the rich sensor information hinders its practical deployment. Recent studies have explored selection strategies that share only a subset of features per frame while striving to keep the performance on par. Nevertheless, the bandwidth requirement still stresses current wireless technologies. To fundamentally ease the tension, we take a proactive approach, exploiting the temporal continuity to identify features that capture environment dynamics, while avoiding repetitive and redundant transmission of static information. By incorporating temporal awareness, agents are empowered to dynamically adapt the sharing quantity according to environment complexity. We instantiate this intuition into an adaptive selection framework, COOPERTRIM, which introduces a novel conformal temporal uncertainty metric to gauge feature relevance, and a data-driven mechanism to dynamically determine the sharing quantity. To evaluate COOPERTRIM, we take semantic segmentation and 3D detection as example tasks. Across multiple open-source cooperative segmentation and detection models, COOPERTRIM achieves up to 80.28% and 72.52% bandwidth reduction respectively while maintaining a comparable accuracy. Relative to other selection strategies, COOPERTRIM also improves IoU by as much as 45.54% with up to 72% less bandwidth. Combined with compression strategies, COOPERTRIM can further reduce bandwidth usage to as low as 1.46% without compromising IoU performance. Qualitative results show COOPERTRIM gracefully adapts to environmental dynamics, localization error, and communication latency, demonstrating flexibility and paving the way for real-world deployment.