Skip to content

Latest commit

 

History

History
50 lines (33 loc) · 3.04 KB

index.md

File metadata and controls

50 lines (33 loc) · 3.04 KB

Cone detection using YOLOv2

Approach

  1. Data collection
    1. Created a dataset of ~300 manually annotated instances of all the 5 colors of traffic cones
  2. Setup my the environment in ubuntu to make ros node
    1. Had a lot of trouble with the setup. Thanks for the heads up about the ROS source and Py3
    2. After trying yolov3 ,darknet, I decided darkflow was the best way
  3. Detection without cones - Getting the basics right
    1. Object detection with darkflow.

    output

    Check output/code here

    1. Object detection in video

    video

  4. Training the model on my "CPU" laptop ( Not my first time )
    1. I trained 200 videos on my laptop in windows in couple of hours but in after changing to ubuntu for this training program, the epoch per hour for these 'image' dataset is so high.
    2. Found pretrained weights but no config file. Morphed many configs and experimented, no use.
  5. Turning to Google colabs
    1. Redone the entire setup again My colab GPU env. Entire darkflow, dataset, models and configurations in the above link. (Note : Not fully functoning)
    2. Got struck training in colab due to format/access errors raised by google drive. Still fiddling with it....

.and so on but unfortunately could not go any further

Future Work

  1. Ill try to get it to work. Darkflow is meant to work for py3 but i discovered it works well for py2 also. So hopefully it doesnt conflict with ROS

Update after the deadline

  1. Editing out all the conflicts Darkflow had with GDrive and hardcoded the classes and model config. It kind of worked.
    1. model - tiny yolo
    2. training- 300 images
    3. epoch -100,200 - 5k , 10k steps
    4. convergence - ~1
    5. GPU- Tesla 180 - colabs
    6. FPS- 4 on gpu and trained on tiny yolo architecture with ~200 epoch- 10k steps for 5 hours on tesla GPU in colab. Checkpoints however werent stored (didnt even prompt an error) after 5000 steps. Testing on the video provided at a speed of 4 FPS on a CPU. And the result looked something like this. cones and whole video is here. It detects blueCones pretty well but not white. Intrestingly I provided more orange samples than the blues