RVSS2019-WS

Repo for the workshop part of the Australian Centre for Robotic Vision Summer School RVSS2019

Data Collection

Use on_robot/collect_data/collect.py for collecting images. Robot control:

• Direction is defined by speed of left and right wheels: ppi.set_velocity(left, right).
• General speed is set in Kd parameter, Ka defines how much speed is changed when robot is turning.
• Press up to go straight: left = right = Kd.
• Press right to turn right: left += Ka; right -= Ka.
• Press left to turn left: left -= Ka; right += Ka.
• When turning in the opposite direction, speed is first reset to Kd.
• Press space to stop: ppi.set_velocity(0, 0).

Images are saved to on_robot/collect_data/data. Image naming example: 000001_18_14.jpg, where 000001 - image id, 18 - speed of the left wheel, 14 - speed of the right wheel.

Data Preparation

1. Put all the collected data in folders dataset1, dataset2, etc (no more than 999 images per folder or there'll be problems with names).

2. Use rename.py to rename all images and put them to processed_data folder. The new naming: 010001_right2.jpg, where first two digits (01) identify dataset folder, las 4 digits (0001) - image id, string (right2) - type of turn.

3. Use split_by_turn.py to split the images to different folders according to the type of turn.

4. Use data_augmentation.py to mirror the images. If 3rd digit in image id is '1' then image was mirrored. (Ex: 030001_right1.jpg - original image; 031001_left1.jpg - mirrored image.)

5. Use delete.py to cut down the number of images (for class balance).

6. Use create_dataset.py to split images into training and validation sets, so that data for each turn is splitted 70/30.

7. Move train_data and val_data folders to on_laptop/dev_data.

Training

• on_laptop/steer_net/steerDS.py contains SteerDataSet class for loading (image, label) dataset instances.
• on_laptop/steer_net/steerNet.py contains steerNet network with following architecture:
• on_laptop/steer_net/train_model.py contains function train_model that performs training and validation steps and saves the model that achieves the best accuracy on validation data.
• Run on_laptop/train.py to train the model. Output directory for model weights and information files is on_laptop/trained_models/{random number}.

Testing

1. Move model file steerNet.py and model weights steerNet.pt to on_robot/deploy.
2. Run on_robot/deploy/deploy0.py.