search

bsridatta / Cone-Detection

Cone Detection using dark flow YOLOv2 for formula student. Custom Dataset from google images
10 stars 6 forks source link

readme

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) Trouble with the ROS source and Py3. They do not work well together.
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](https://github.com/bsridatta/kthfsdv-Perception-Interview/blob/master/darkflow.png)

  Check [output/code](https://github.com/bsridatta/kthfsdv-Perception-Interview/blob/master/py2%20Darkflow.ipynb) here

  2) Object detection in video 

  ![video]("https://github.com/bsridatta/kthfsdv-Perception-Interview/blob/master/video.avi")

4) ##### Training the model on my "CPU" (laptop) 

    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](https://github.com/melfm/dukecone) 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. 3) Modified darkflow's util functions

Modification

1) Editing out all the conflicts Darkflow had with GDrive and hardcoded the classes and model config. It kind of works. 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 cpu 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