- To detect with onboard computer of UAV, there is: Light U-net for wildfire segmentation
- Steps to Run
- Change the dataset folder in data to load in the prepared images and masks to check.
- Check whether the Network strucure is okay enough, this model is based on Resnet18 Encoder-Decoder, add layers for more complex segmentation missions.
- Find appropriate learning rate based on Leslie method with lrfinder, there is an appliable package could simply work it.
- Ready to train the model, traing on train
- Test on images with imgtest and videos with videotest. The opencv is recommonded to load the video and image, but ther is convert part happen in the
.pyfiles.
-
.pyfiles are relying on the blocks and tools in deving -
util.py: Some functions are stored, such as: saving model at every iteration; save entire model; saving predicted result; saving accuracy's and loss's plots; plot image, pred, mask in a figure ... -
Cross Entropy Loss commonly, use Sum of the Squared Residuals to determine how well the Neural network fits the data.
$$SSR = \sum_{i = 1}^{n = 3}{(Observed_i - Predicted_i)^2}$$