fire-detection-from-images

fire-detection-from-images copied to clipboard

Use smaller yolov5

1

Have been training with `--img 416` but images are approx 285x180 pixels, meaning I have been oversized. This means the generated anchor boxes are probably also oversized

robmarkcole

Experiment with augmentation more

1

Getting pretty unsatisfactory performance. The article below highlights that flames are a difficult target: `wildfires have one of the most complicated features to obtain through CV, including flame height, flame...

robmarkcole

https://towardsdatascience.com/tools-to-annotate-and-improve-computer-vision-datasets-f9b99cdb0e04

robmarkcole

https://github.com/jveitchmichaelis/edgetpu-yolo

robmarkcole

https://pymde.org/

robmarkcole

Checkout ice vision-dashboard

1

https://github.com/fstroth/icevision_dashboard Could assist with filtering small objects

robmarkcole

Investigate anchor boxes

1

Raised by Farid on discord: Try changing the anchor boxes sizes to test if it can improv model accuracy - Read: https://medium.com/@andersasac/anchor-boxes-the-key-to-quality-object-detection-ddf9d612d4f9 - Read: https://blog.roboflow.com/what-is-an-anchor-box/ Some takeaways: Improving Anchor Box...

robmarkcole

Simple classification example using pytorch(icevison)

1

Hello, would you be able to provide an example which takes an image as input and returns is there is fire and where? Using pre-trained models.

TheGreyDiamond

Try retinanet

1

https://github.com/airctic/icevision/pull/538 https://airctic.com/retinanet/ On master. In colab: `pip install git+git://github.com/airctic/icevision.git#egg=icevision[all] --upgrade`

robmarkcole

cocometric reported by retinanet (icevision) is comparable to mAP_0.5 from yolov5 training script. In icevision use `print_summary = True` to get all metrics 

robmarkcole