QuantV2X: A Fully Quantized Multi-Agent System for Cooperative Perception
The official research code for an ECCV 2026 paper on quantizing every part of a multi-agent cooperative perception system for autonomous vehicles, aimed at cutting inference and communication cost with minimal loss of perception accuracy.
QuantV2X is the reference implementation accompanying a paper addressing latency and performance problems in cooperative perception, where multiple vehicles or roadside units share sensor information for real-world, resource-constrained deployment. It introduces a fully quantized multi-agent system, including a codebook learning stage and a post-training quantization pipeline, intended to preserve perception performance while reducing the cost of model inference and inter-agent communication.
The codebase supports full-precision baseline training and inference on the V2X-Real dataset, as well as on the OPV2V and DAIR-V2X datasets, and includes a TensorRT deployment pipeline. It is part of a broader line of work on cooperative perception alongside two related projects, CooPre and TurboTrain, from the same research group.
It is written in Python and was presented as ICCV 2025 tutorial material and an ICCV 2025 workshop oral presentation before its ECCV 2026 acceptance.