YOLOP_RESA

[TOC]

1, Introduce

---

In this project, we aim to make a Lane_detection System for Self Driving Car with Instance Segmentation Approach. This system can detect not only weather each pixel is a part of Lanes but which Lane it belongs to.

2, Structure

---

We use Instance Segmentaion as our main approach, We apply CNN to deal with our task. Our CNN Net consist of two part:

• YOLOP net (CSPDarknet with SPP structure)
• RESANET (RESA and BUSDecoder)

3, Training

---

▲Preparing

Our Training data need to be labelled in the form of Tusimple Dataset ,which means your dataset must consist of two part: your gt_image and json file.

▲Step 1

With the labelled data, your need to put your dataset folder in ./data and create another three .json file -- test_label.json and valset.json

data/your_data_folder 
    |-gt_image/ 
    |-JsonFile_label.json
    |-test_label.json
    |-valset.json

▲Step 2 prepare testing set

then, cut (or copy ) some lines in JsonFile_label and paste them in test_label.json and valset.json

▲Step 3 generating segmentation labels

python tools/generate_seg_tusimple.py --root ./data/your_data_folder

▲Step 4 modify config python file

1.set your training path

go to ./lib/config/resa_tusimple.py, find "dataset_path" and "test_json_file" and change their content with './data/your_data_folder' and './data/your_data_folder/test_label.json' respectively.

▲Step 5 training

Start training using following comment:

python tools/train.py --view --cpu 0 --load_from your/pretrained/weight/path

4,testing

python tools/train.py --view -- validate --cpu 0 --load_from your/pretrained/weight/path

---