Aerial_drone_image_segmentation

Dataset Resource:

• https://www.tugraz.at/index.php?id=22387

Dataset Overview

The Semantic Drone Dataset focuses on semantic understanding of urban scenes for increasing the safety of autonomous drone flight and landing procedures. The imagery depicts more than 20 houses from nadir (bird's eye) view acquired at an altitude of 5 to 30 meters above ground. A high resolution camera was used to acquire images at a size of 6000x4000px (24Mpx). The training set contains 400 publicly available images.

---

Task	Colab
Colab Notebook for Aerial Semantic Segmentation : Running yourself

Aerial Semantic Segmentation Drone Sample Images with mask

---

---

What is semantic segmentation ?

• Source: https://divamgupta.com/image-segmentation/2019/06/06/deep-learning-semantic-segmentation-keras.html

• Semantic image segmentation is the task of classifying each pixel in an image from a predefined set of classes.

---

In the following example, different entities are classified.

---

In the above example, the pixels belonging to the bed are classified in the class “bed”, the pixels corresponding to the walls are labeled as “wall”, etc.

In particular, our goal is to take an image of size W x H x 3 and generate a W x H matrix containing the predicted class ID’s corresponding to all the pixels.

---

---

Usually, in an image with various entities, we want to know which pixel belongs to which entity, For example in an outdoor image, we can segment the sky, ground, trees, people, etc.

---

Data block to feed the model

• Created using fastai datablocks API.

data = (SegmentationItemList.from_folder(path=path/'original_images')  # Location from path
        .split_by_rand_pct(0.1)                          # Split for train and validation set
        .label_from_func(get_y_fn, classes=codes)      # Label from a above defined function
        .transform(get_transforms(), size=src, tfm_y=True)   # If you want to apply any image Transform
        .databunch(bs=4)                                   # Batch size  please decrese batch size if cuda out of memory
        .normalize(imagenet_stats))            # Normalise with imagenet stats
data.show_batch(rows=3)

---

---

Model | unet_learner

Fastai's unet_learner

• Source Fast.ai

• This module builds a dynamic U-Net from any backbone pretrained on ImageNet, automatically inferring the intermediate sizes.

• This is the original U-Net. The difference here is that the left part is a pretrained model.

• This U-Net will sit on top of an encoder ( that can be a pretrained model -- eg. resnet50 ) and with a final output of num_classes.

void_code = -1
def accuracy_mask(input, target):
    target = target.squeeze(1)
    mask = target != void_code
    return (input.argmax(dim=1)[mask]==target[mask]).float().mean()

arch =  pretrainedmodels.__dict__["resnet50"](num_classes=1000,pretrained="imagenet")

learn = unet_learner(data, # DatBunch
                     arch, # Backbone pretrained arch
                     metrics = [metrics], # metrics
                     wd = wd, bottle=True, # weight decay
                     model_dir = '/kaggle/working/') # model directory to save

Training :

• Learning rate : Used fastai's lr_find() function to find an optimal learning rate.

learn.lr_find()
learn.recoder.plot()

callbacks = SaveModelCallback(learn, monitor = 'accuracy_mask', every = 'improvement', name = 'best_model' )
learn.fit_one_cycle(10, slice(lr), pct_start = 0.8, callbacks = [callbacks] )

---

epoch	train_loss	valid_loss	drone_accuracy_mask	time
0	2.106114	1.835669	0.483686	00:16
1	1.697477	1.331731	0.618320	00:12
2	1.462504	1.227808	0.659259	00:13
3	1.455247	1.147135	0.691440	00:12
4	1.336305	1.115214	0.691359	00:12
5	1.247270	1.168395	0.669359	00:12
6	1.310875	1.181834	0.672961	00:13
7	1.196860	1.115905	0.708513	00:12
8	1.131353	0.888599	0.778729	00:12
9	0.983422	0.799664	0.851884	00:12
10	0.973422	0.779664	0.881884	00:12
11	0.953422	0.749664	0.911884	00:12

---

Results |

Intial dynamic unet on top of an encoder ( resnet34 pretrained = 'imagenet' ), trained for 10 epochs gave an accuracy of 91.1% .

learn.show_results(rows=3, figsize=(12,16))