1. Generate the Inflated Occupancy Annotations as described here:
- Set
only_generate_dataset=True
in both projects/configs/inflated/action_condition_GMO.py and projects/configs/inflated/action_condition_MMO.py - Run the training and evaluation scripts using these configurations. Since the model's forward process is not executed, this will not consume GPU memory:
CONFIG=projects/configs/inflated/action_condition_GMO.py
GPU_NUM=8 # 8 GPUs(Processes) consume ~20 hours
# Generate training data
./tools/dist_train.sh ${CONFIG} ${GPU_NUM}
# Generate validation data
CKPT=work_dirs/action_condition_GMO/epoch_1.pth
./tools/dist_test.sh ${CONFIG} ${CKPT} ${GPU_NUM}
2. Train
CONFIG=path/to/config.py
GPU_NUM=8
./tools/dist_train.sh ${CONFIG} ${GPU_NUM}
Configuration Files:
- inflated/action_condition_GMO.py: Predicts General Moveable Objects (GMO) using the inflated occupancy annotations as defined in Cam4DOcc.
- inflated/action_condition_MMO.py: Predicts Multi-class Moveable Objects (MMO) using the inflated occupancy annotations defined in Cam4DOcc.
- fine_grained/action_condition_GMO.py: Predicts General Moveable Objects (GMO) using the fine-grained occupancy annotations defined in nuScenes-Occupancy.
- fine_grained/action_condition_GMO_GSO.py: Predicts General Moveable Objects (GMO) and General Static Objects (GSO) using the fine-grained occupancy annotations defined in nuScenes-Occupancy.
- fine_grained/action_condition_MMO_MSO.py: Predicts Multi-class Moveable Objects (MMO) and Multi-class Static Objects (MSO) using the fine-grained occupancy annotations defined in nuScenes-Occupancy.
3. Evaluate
CONFIG=path/to/config.py
CKPT=work_dirs/config_file_name/epoch_24.pth
GPU_NUM=8
./tools/dist_test.sh ${CONFIG} ${CKPT} ${GPU_NUM}