kaiopen / tabdet

Travelable area boundary detection models.

Travelable Area Boundary Detection

中文

Python Environments

• tqdm
• NumPy
• PyTorch
• KaiTorch
• TABKit

Preparing

The prepared will be save in the directory assigned by the item preprocess.dst in the configuration file. Run the command:

python tools/preprocess.py --config=<configuration file> --split=<split>

For example,

python tools/preprocess.py --config=TAB_Pillar256x512x20_UNet.yaml --split=train

TRAINING

The training will use the GPU devices as much as can be gotten. You can limit and assign the device(s) via CUDA_VISIBLE_DEVICES. Checkpoints will be save in ./checkpoints/.

torchrun --nproc_per_node=2 tools/train_ddp.py --config=TAB_Pillar256x512x20_UNet.yaml --split=train --batch_size=16 --num_worker=10 --end_epoch=200

EVALUATION

A loss curve figure will be generated and saved in ./eval/. We have not done evaluation when training. You can modify the tools/train_ddp.py and try to do evaluate during training.

python tools/eval.py --config=TAB_Pillar256x512x20_UNet.yaml --batch_size=64 --num_worker=10

VISUALIZATION

Prediction results will be visualized and save in ./vis.

python tools/vis_tab.py --config=TAB_Pillar256x512x20_UNet.yaml --split=test --batch_size=64 --num_worker=10 --checkpoint=199

TEST

Assess the well-trained models. Results will be saved in ./results/.

python tools/test.py --config=TAB_Pillar256x512x20_UNet.yaml --split=test --batch_size=64 --num_worker=10 --checkpoint=199

Models

Backbone	Straight-going side				Turning				Ignoring semantics
	mF_p	F_0.3	F_0.5	F_0.8	mF_p	F_0.3	F_0.5	F_0.8	mF_p	F_0.3	F_0.5	F_0.8
UNet	0.55	0.50	0.41	0.07	0.71	0.68	0.63	0.41	0.75	0.70	0.67	0.32
HRNet-w18	0.69	0.66	0.63	0.56	0.73	0.75	0.69	0.54	0.82	0.83	0.77	0.62
DeepLabV3+	0.62	0.57	0.49	0.35	0.71	0.67	0.61	0.43	0.78	0.73	0.67	0.46

NOTE: The checkpoint of the DeepLabV3+ has not been uploaded because it is larger than 2 GB. You can rise up an issue and leave your email address. The checkpoint will be sent to you as soon as possible.

You can download well-trained checkpoints and move them into checkpoint directories such as ./checkpoints/TAB_Pillar256x512x20_UNet.yaml/.