This folder contains our code for the second part of submission which is Segmentation of Images. For segmentation task we UNET model, where we use 4 encoding layers and 4 decoding layers.

In this section, our objective is to train the U-Net model on train set images provides and evaluate the model's performance on validation set.

U-Net : CNN based model for semantic segmentation of objects in 22nd Frame.

• 4 encoding layers and 4 decoding layers

• 1 bottle neck layer

• trained on 22 x 1000 images of training set

• trained for 10, 20 and 50 epochs.

• learning rate : 1e-4

• optimiser : Adam

• loss : Cross Entropy

• evaluated on 1000 images of validation set.

• evaluation metric : Jaccard index

Dataset used in this section is:

• 1,000 labeled training videos with 22 frames each, Each video folder has masks.npy which contains masks of 22 all 22 frames in the video.

Thus, we inflated all the frames and their respective masks in the "SegmentationDataSet" class, and shuffled all the images. Now total 22 X 1000 images, i.e. 22000 frames(images) are used to train model.

• 1,000 labeled validation videos with 22 frames each, this is used for validating the images performance.

1. In main function of main.py, specify the path of train set and validation set.

python main.py

2. After the running the main.py, it will save trained U-Net model as unet.py.
3. Now to check the performance of trained unet model, on validation set, run the Checking_Jaccard_index.ipynb file.
4. In the 3rd cell of .ipynb file specify the path of "UNET trained" model unet.pt and "validation set".
5. Run all the cells of the Checking_Jaccard_index.ipynb file, and final jaccard score on validation set can be derived.