/Indian-Driving-Dataset-Semantic-Segmentation

Primary LanguageJupyter NotebookMIT LicenseMIT

A Deep Learning project on Multi-class Semantic Segmentation on Indian Driving Dataset

About Dataset

I have used dataset which is available at kaggle IDD dataset

Dataset provides image resolution of (width:1920, height:1080):

  1. Training Images: 5966
  2. Validation Images: 1016

I have used Validation images as Test Set for evaluation purpose and For training i have splitted Training images to train(4950) and Val set(1016)

The label image consists of 26 classes.

For this project, i have customized the number of classes:

  1. There are 2 main classes:

    • Road (Drivable)
    • Moving objects (Living things and Vehicles category as per this offical website)
    • Background (Rest of the categories as background which includes Non Drivable, Road Side Objects, Sky, Far Objects)

After mapping it to 3 categories, label image looks like:

image

Script to customize number of classes for IDD dataset is available at Scripts/DatasetPreparation/prepare_IDD_dataset.ipynb

Training

  • For training i have resized images to 512 X 512

    image

  • I have trained various models on this dataset with different techniques. All training scripts are present in Scripts/TrainingScripts/

    • Deeplabv3p with MobilenetV2 with pretrained Cityscapes weights
    • Deeplabv3p with Xception with pretrained Cityscapes weights
    • Unet with MobilenetV2 with pretained encoder
    • Unet with EfficientNetB0 with pretained encoder
    • Unet with EfficientNetB1 with pretained encoder
    • Unet with EfficientNetB6 with pretained encoder
    • Unet with PSPNet with pretained encoder
    • Unet with FPN with pretained encoder
    • Unet with LinkNet with pretained encoder
    • Deeplabv3p with MobilenetV2(alpha 1.0) - custom model with pretrained encoder
    • Deeplabv3p with MobilenetV2(alpha 0.35) - custom model with pretrained encoder and decoder upsampled by 2
    • Deeplabv3p with MobilenetV2(alpha 0.35) - custom model with pretrained encoder and decoder upsampled by 4
    • Deeplabv3p with MobilenetV2(alpha 0.35) - custom model with pretrained encoder and decoder upsampled by 4, applied DepthwiseSeparable Convolution in decoder
    • Deeplabv3p with MobilenetV2(alpha 0.3) - custom model trained from scratch

Evaluation

I have Evaluated models results on datasets:

Results

Model Images Count mIoU Precision Recall F1-score (Dice-score) F2-score
IDD_Unet_efficientnetB6_alpha1_102epoch_IDD 1016 0.935507686 0.966729734 0.964703715 0.964771986 0.964548665
IDD_Unet_efficientnetB1_alpha1_169epoch_IDD 1016 0.930382308 0.962567004 0.963652336 0.961783635 0.962565754
IDD_Unet_efficientnetB0_alpha1_172epoch_IDD 1016 0.929789305 0.964161625 0.961232964 0.961725158 0.961219339
IDD_Unet_mobilenetV2_alpha1_125epoch_IDD 1016 0.924768612 0.96202396 0.957881042 0.958814571 0.958030642
IDD_Dv3p_mobilenetV2_upsampleby2_custom_alpha0_35_168epoch_IDD 1016 0.923801275 0.961042069 0.957990114 0.958356378 0.957920497
IDD_Linknet_mobilenetV2_alpha1_100epoch_IDD 1016 0.919378631 0.959779189 0.953338128 0.95566278 0.954095486
IDD_Unet_mobilenetV2_alpha0_35_150epoch_IDD 1016 0.917968092 0.961496856 0.950309282 0.954694436 0.951842498
IDD_Dv3p_mobilenetV2_upsampleby4_1X1_DSC_custom_alpha0_35_340epoch_IDD 1016 0.912327724 0.955982275 0.948532736 0.951089052 0.949336086
IDD_Dv3p_mobilenetV2_upsampleby4_custom_alpha0_35_111epoch_IDD 1016 0.911748393 0.955098572 0.949369409 0.951108694 0.949833539
IDD_PSPnet_mobilenetV2_alpha1_125epoch_IDD 1016 0.889653239 0.941406104 0.937448567 0.937932286 0.937336957
IDD_Unet_efficientnetB1_alpha1_169epoch_CS 500 0.736945095 0.813979215 0.878560476 0.834813002 0.857184156
IDD_Unet_efficientnetB6_alpha1_102epoch_CS 500 0.72419415 0.803299049 0.868500158 0.823064889 0.845677884
IDD_Linknet_mobilenetV2_alpha1_100epoch_CS 500 0.707557416 0.810975559 0.836996694 0.815088204 0.82578831
IDD_mobilenetV2_apha1_725epoch_CS 500 0.704201384 0.820824768 0.834737747 0.817251204 0.824993245
IDD_Unet_mobilenetV2_alpha1_125epoch_CS 500 0.703853585 0.792599002 0.847816488 0.80684931 0.826419052
IDD_Unet_efficientnetB0_alpha1_172epoch_CS 500 0.702946435 0.775713217 0.84451824 0.796338827 0.81949006
IDD_mobilenetV2_apha1_350epoch_CS 484 0.684523423 0.806061196 0.811062906 0.794792795 0.800980755
IDD_mobilenetV2_apha1_350epoch_BDD 1000 0.68348494 0.782480068 0.828049944 0.783499953 0.80064466
IDD_Unet_efficientnetB6_alpha1_102epoch_BDD 1000 0.648662205 0.774714055 0.803258877 0.750631798 0.766699096