Short demo in GIF here
Codes will be released upon paper acceptance
Download to ./datasets
You can download all the KITTI raw datasets with a single command using the repository available at this repo.
Integrated checkpoints:
_C.CKPT.ORI_CKPT + _C.CKPT.TRAJ_PRED_CKPT
export PYTHONPATH=$pwd:PYTHONPATH
python ./tools/main.py
_C.TRAIN.TASKS = "pred"
Run on KITTI:
_C.EVAL.DATASET_NAME == 'kitti_raw'
Visualize:
_C.VISUALIZE = True
Saved path: <_C.EVAL.OUTPUT_DIR>/saved_frames
Use tools/frame2video.py
to render video
Test Speed:
_C.EVAL.SPEED_TEST = Tr
_C.TRAIN.TASK = 'pred'
Evaluate segmentation on city dataset:
`_C.EVAL.TASK = 'seg'`
`_C.EVAL.SEG_DATASET = 'city'`
Evaluate depth:
_C.EVAL.TASK = 'depth'
Checkpoints output directory:
<_C.TRAIN.OUTPUT_DIR>/<_C_TRAIN.TASKS>/checkpoints
Train from scratch:
_C. TRAIN.START_NEW = True
Train every module with checkpoints loaded:
_C. TRAIN.START_NEW = False
_C.TRAIN.ALL = True
Train individual backbone + neck + task head;
Road segmentation: _C.TRAIN.TASKS = 'seg'
Detection: _C.TRAIN.TASKS = 'det'
Train backbone + depth:
_C.TRAIN.TASKS = 'depth'
Other training:
Backbone + Neck + Detection Head + Segmentation Head:
_C.TRAIN.TASKS = 'det-seg'