Week 2 - Amansangram/darknet_aman GitHub Wiki

Week-2 (25/05/2020 – 31/05/2020)

Tuesday (26/05/2020)

Peer teaching went well. Got good reviews from teacher and suggestions to improve the presentation skills. Now going back to the project. I am using the lane follower code provided by our tutors, and will be trying to integrate either YOLO or CNN.

Wednesday (27/05/2020)

As the YOLO detection is not working properly, I have decided to explore CNN a little more. The code I am using is giving me high accuracy.

  • After training the datasets, the model will be stored in a pickle file.
  • This pickle file will be loaded in the test file which will using live webcam video. Now I have to call the results from this detection in the lane follower code.

Thursday (28/05/2020)

  • Doubt session was taken by the teacher on zoom. It was pointed out that doing altering the threshold value in YOLO may help improve the detection.
  • I changed the threshold and the results improved. Only one of my class had training good enough to have a probability of more than 60%, while others were below 40%.
  • Decided to do training again with new and better datasets for longer period of time.

Friday (29/05/2020)

The detection is working fine, but unable to use the results in the lane following code. ROS doesn't recognizance the pickle file because ROS operates in Python3 while .pickle format is supported by Pytho2 only. New approach has to be used.

  • Used new ROS package called cnn_bridge which will help in publishing the detection results. Further I can subscribe to those results in the lane follower code and use them. (cnn_bridge http://wiki.ros.org/cnn_bridge)
  • This package publishes:
  1. detection
  2. segmentation.
  • The package has following parameters:
  1. source
  2. ogdir
  3. metadata_source
  4. mode
  5. input_tensor
  6. output_tensor
  7. display
  8. save_mode 9.cpu
  9. gpu_percent
  • The package need modification in launch files.

Sunday (31/05/2020)

  • Have been struggling to removes errors and modify launch files, but the package still shows problems.
  • As decided earlier, I created new data using the duckie. Now I have three classes with each one have more than 50 images each.
  • I was unable to train the data now because of the issues with github. Will be training YOLO next week.