MapKD: Unlocking Prior Knowledge with Multi-Level Cross - Modal Distillation for Map-aware Driving Perception

Replication package access:  https://github.com/2004yan/MapKD_new. Data description: We adopt the publicly available nuScenes as our dataset. Code description:  torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 Simulation software description: Configuring Python 3.8 on Ubuntu 20.04 Experiment design description: In the first stage, the teacher and coach model is jointly pretrained.In the second stage, the student model is trained from scratch under the proposed Teacher-Coach-Student pipeline. The teacher and the coach are kept frozen, while providing intermediate guidance at the feature and output level. The student is supervised with both hard labels and soft logits. We train our model on the nuScenes dataset , using a BEV range of 60 m × 30 m. The voxel resolution is set to 0.15 m. Input images are resized to 128 × 352. We train for 30 epochs with a batch size of 8, using 4 GPUs and 20 workers.We use the Adam optimizer with a learning rate of , weight decay of , and gradient clipping at 5.0.