CoDynTrust: Robust Asynchronous Collaborative Perception via Dynamic Feature Trust Modulus

ICRA 2025 paper: "CoDynTrust: Robust Asynchronous Collaborative Perception via Dynamic Feature Trust Modulus" code

arXiv | Demo | Code

Because I'm lazy, I won't write the readme in detail. The code is a reproducible version. Please refer to its code for training and inference. For environment configuration, please refer to CoBEVFlow's configuration, because CoDynTrust is based on CoBEVFlow's code baseline. In fact, CoDynTrust contains two stages of training. The first stage is to obtain an uncertainty evaluation model (that is, a single-agent model, such as pointpillars as a backbone single-agent model). I improved this evaluation model using CoAlign's baseline. For the relevant code, you can refer to the repo INSTINCT of my other work, which contains this part of the code. The single-agent model with 'uncertainty' in the name has the relevant code. The second stage of training is to freeze the parameters of the first stage and train the subsequent collaborative fusion part. At the same time, the detection head will also be decoupled. The relevant code is in this CoDynTrust repository. dair_cobevflow_w_uncertainty.yaml is required to complete the training requirements on the DAIRV2X data set.

Acknowledgement

Many thanks to these excellent open source projects:

• OpenCOOD
• Where2comm
• CoAlign
• CoBEVFlow