Todo

• Make labelling tool for clips
• Make clip visualiser tool
• Label ~10 clips to make a small evaluation set

Bulletin Board

• 8/4/2021 - Daniel: Please make sure you have a models/ folder if you want to run the detection classifier.
• 7/6/2021 - Dean: Make labelling tool and labelled clip visualiser.
• 6/6/2021 - Dean: Moved datasets to data/ folder for labelling (please do this locally).
• 4/6/2021 - Welcome to the start of a long nightmare. As everyone's detection algorithm is poor, please use the labels for now for the bounding boxes until a better detection algorithm is created.

Usage

data/labeller.py

Manually label clips starting at clip_number/frame_number from source_clips_path. Outputs copies of images to paths_to_outputs (e.g. see evaluation/clips). With no arguments, default options will be used.

python labeller.py [ --clip | -c ] clip_number
                   [ --frame | -f ] frame_number
                   [ --src | -s ] source_clips_path
                   [ --dst | -d ] path_to_outputs

Example

$ cd data
$ python labeller.py --clip 2 --frame 3

data/watch_clip.py

python watch_clip.py  [ --clip | -c ] clip_number
                      [ --src | -s ] source_clips_path

Example

$ cd data
$ python watch_clip.py --clip 1

Setup

Please download and extract the zip files in both:

1. VELOCITY ESTIMATION CHALLENGE
2. LANE DETECTION CHALLENGE

From TuSimple/tusimple-benchmark#3 by Kivinju (on github)

Then extract the zip files, there now should be 5 folders located at the root of this folder:

1. train_set
2. test_set
3. benchmark_velocity_train
4. benchmark_velocity_test
5. benchmark_velocity_supp

This is to keep things consistent. Also, please ensure no generated files e.g. generated images are stored on the github. Add the folders to .gitignore ^_^

---

Helpful Links: https://unsw-my.sharepoint.com/personal/z5195702_ad_unsw_edu_au/_layouts/15/onedrive.aspx?id=%2Fpersonal%2Fz5195702%5Fad%5Funsw%5Fedu%5Fau%2FDocuments%2FLane%20Detection%20Data%20Subset&originalPath=aHR0cHM6Ly91bnN3LW15LnNoYXJlcG9pbnQuY29tLzpmOi9nL3BlcnNvbmFsL3o1MTk1NzAyX2FkX3Vuc3dfZWR1X2F1L0VucVlpRnhULTRaTWlsbS04b1ZsVXVzQk1qOGg0cExDbFpNd2Z0VEtNbVRxZkE%5FcnRpbWU9ek02MVo1bjQyRWc

For train_set as train_set is ~10gb. However, do not use this for the final submission.... (You can use this to start early and dl giant dataset overnight)