Train and Test Custom Data

This page explains how to train and test your own custom data with LSTR.

We provide sample images and annotations in ./raws, just make your dataset the same as them.

0. Before you start

Clone this custom branch

git clone https://github.com/liuruijin17/LSTR.git -b custom

and follow the README to install LSTR.

1. Prepare your own dataset

Step 1 Prepare your own dataset with images and labels first. For labeling images, you can use tools like Labelme or CVAT.

Step 2 Then, you should write some scripts to transfer your annotations into .txt files and make sure:

1. each image (.jpg) and its annotation file (.txt) has the same name;
2. in the .txt file, each row store the set of points for one lane;
3. for each row, points are stored by x1 y1 x2 y2...

If aforementioned descriptions are still hard to understand, see .txt files in ./raws.

Step 3 Split your data into train and test by putting training images into ./raws/train_images and their corresponding annotation .txt files into ./raws/train_labels So does for testing data.

2. Train your own dataset

python train.py LSTR

3. Test your own dataset

python test.py LSTR --modality eval --split testing --testiter 500000

Since the provided sample images and annotations in ./raws are directly transformed from TuSimple, you can run above test command to get a F1 result first. If everything is running correctly, you would see 0.79 F1 result.