DAIR-V2X is the first large-scale and real-world vehicle-infrastructure cooperative 3D object detection dataset. This dataset includes the DAIR-V2X-C, which has the cooperative view. We train and evaluate the models on DAIR-V2X dataset. For downloading DAIR-V2X dataset, please refer to the guidelines in DAIR-V2X.
We construct the frame pairs to generate the json files for FFNet training and evaluation.
- flow_data_info_train_2.json: frame pairs constructing from DAIR-V2X-C to simulate the different latency for training, including k=1,2
- flow_data_info_val_n.json: frame pairs constructing from DAIR-V2X-C to simulate the different latency for evaluation. n=1,2,3,4,5, corresponding to the 100ms, 200ms, 300, 400ms and 500ms latency, respectively.
- example_flow_data_info_train_2.json: frame pairs constructing from the example dataset to simulate the different latency for training, including k=1,2
- example_flow_data_info_val_n.json: frame pairs constructing from the example dataset to simulate the different latency for evaluation. n=1,2,3,4,5, corresponding to the 100ms to 500ms latency.
The json files are used for spliting the dataset into train/val/test parts.
Please refer to the split_data for the latest updates.
We use the DAIR-V2X-C-Example to illustrate how we preprocess the dataset for our experiment. For the convenience of overseas users, we provide the original DAIR-V2X-Example dataset here. We provide the preprocessed DAIR-V2X-C-Example dataset here.
# Preprocess the dair-v2x-c dataset
python ./data/dair-v2x/preprocess.py --source-root ./data/dair-v2x/DAIR-V2X-Examples/cooperative-vehicle-infrastructure
We have provided the frame pair files in flow_data_jsons. You can generate your frame pairs with the provided example script.
# Preprocess the dair-v2x-c dataset
python ./data/dair-v2x/frame_pair_generation.py --source-root ./data/dair-v2x/DAIR-V2X-Examples/cooperative-vehicle-infrastructure