SOLOv2_Road-Markings

1. Converts the format of VGG Image Annotator (VIA) to the COCO format.

Classes:straight arrow, left arrow, right arrow, straight left arrow, straight right arrow, pedestrian crossing, special lane

2. SOLOv2 environment

• Operating System: Ubuntu 20.04.4
• GPU: NVIDIA GeForce RTX3090
• CUDA 11.1
• pytorch 1.8.0
• torchvision 0.9.0
• python 3.7.13

3. Create my_dataset.py

Create a python file for the classes of the custom dataset in mmdet/datasets

from .coco import CocoDataset
from .registry import DATASETS

@DATASETS.register_module
class MyDataset(CocoDataset):
    CLASSES = ['straight arrow', 'left arrow', 
    'right arrow', 'straight left arrow', 'straight right arrow', 
    'pedestrian crossing', 'special lane']

Add the dataset in the mmdet/datasets/__init__.py

4. Modify solov2_r101_fpn_8gpu_3x.py

Backbone: ResNet101+FPN

• Download pretrained model: https://github.com/pytorch/vision/blob/d585f86d94f07a3bc083e48c6534d93a409cbcb2/torchvision/models/resnet.py#L312
• Modify classes number num_classes: classes number + background
• Modify number of grids num_grids:[80, 72, 64, 48, 32] (if the image size is larger)
• Modify the dataset settings dataset_type data_root
• Modify train/val/test images and annotations path ann_file img_prefix
• Modify runtime settings total_epochs work_dir (the path of the trained weights)

5. Training

  python tools/train.py configs/solov2/solov2_r101_fpn_8gpu_3x.py

6. Evaluation

  python tools/test_ins.py configs/solov2/solov2_r101_fpn_8gpu_3x.py weights/homo_model_2/epoch_100.pth --show --out results_solo.pkl --eval segm

7. Visulization

The class_namesshould be modified.

  python tools/test_ins_vis.py configs/solov2/solov2_r101_fpn_8gpu_3x.py weights/homo_model_2/latest.pth --show --save_dir  work_dirs/val_homo_2data

8. Result

Train with the bird's eye view image model

Front View

Bird's eye view