Late Fusion consists of a 2-stream encoder in which the RGB image and the LiDAR BEV inputs are processed independently of each other. These features are then combined via element-wise summation and passed to waypoint prediction network. This is equivalent to removing the transformer modules from TransFuser.

CUDA_VISIBLE_DEVICES=<gpu_id> python3 train.py --id late_fusion --batch_size 128

Update leaderboard/scripts/run_evaluation.sh to include the following.

export ROUTES=leaderboard/data/evaluation_routes/routes_town05_long.xml
export TEAM_AGENT=leaderboard/team_code/late_fusion_agent.py
export TEAM_CONFIG=model_ckpt/late_fusion
export CHECKPOINT_ENDPOINT=results/late_fusion_result.json
export SCENARIOS=leaderboard/data/scenarios/town05_all_scenarios.json