-
- Created a dataset of ~300 manually annotated instances of all the 5 colors of traffic cones
-
- Had a lot of trouble with the setup. Thanks for the heads up about the ROS source and Py3
- After trying yolov3 ,darknet, I decided darkflow was the best way
-
- Object detection with darkflow.
Check output/code here
- Object detection in video
-
- 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.
- Found pretrained weights but no config file. Morphed many configs and experimented, no use.
-
- Redone the entire setup again My colab GPU env. Entire darkflow, dataset, models and configurations in the above link. (Note : Not fully functoning)
.and so on but unfortunately could not go any further
- 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
- Editing out all the conflicts Darkflow had with GDrive and hardcoded the classes and model config. It kind of worked.
- model - tiny yolo
- training- 300 images
- epoch -100,200 - 5k , 10k steps
- convergence - ~1
- GPU- Tesla 180 - colabs
- 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. and whole video is . It detects blueCones pretty well but not white. Intrestingly I provided more orange samples than the blues