Road-Lane-Instance-Segmentation-PyTorch

Road lane instance segmentation with PyTorch.

• SegNet, ENet with discriminative loss.
• Lane clustered with DBSCAN.
• Trained from tuSimple dataset.
• ROS(Robot Operating System) inference node (20Hz).

ENet result

SegNet result

ROS

$ python2 ros_lane_detect.py --model-path model_best_enet.pth

Train

$ mkdir logs
$ tensorboard --logdir=logs/ &
$ python3 train.py --train-path /tuSimple/train_set/ --epoch 100 --batch-size 16 --lr 0.0001 --img-size 224 224

Dataset

Downloads: tuSimple dataset

Load Dataset

train_path = '/data/tuSimple/train_set/'
train_dataset = tuSimpleDataset(train_path, size=SIZE)
train_dataloader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=16)

Model

ENet summary

----------------------------------------------------------------
Total params: 686,058
Trainable params: 686,058
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 153326.17
Params size (MB): 2.62
Estimated Total Size (MB): 153329.36
----------------------------------------------------------------

SegNet summary

----------------------------------------------------------------
Total params: 29,447,047
Trainable params: 29,447,047
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.57
Forward/backward pass size (MB): 688.68
Params size (MB): 112.33
Estimated Total Size (MB): 801.59
----------------------------------------------------------------

References

https://github.com/nyoki-mtl/pytorch-discriminative-loss
Paper: Semantic Instance Segmentation with a Discriminative Loss Function